When the storms were over and damages started being assessed, NIPSCO, the utility company in the area, published estimated restoration times for Northwest Indiana, and every city on the list got a date and a time. Gary got a word. One resident, watching the list for his cousin, put it this way: “We’re nowhere on the list. It just says ‘undetermined’ for Gary.”

If you have ever waited on anything from an institution, you know the difference between a bad date and no date. A bad date is still a date. You can plan around it. You can decide whether to throw out the milk, whether to send the kids to your sister’s in Chicago, whether it is worth opening the shop on Thursday. Undetermined is not information. Undetermined says keep waiting, and stop asking.

Mayor Eddie Melton said the omission caused a lot of confusion and anger. Bridget Sanders, a resident, told CBS: “We are the last ones on the totem pole to get anything done.”

I want to be careful here, because the easy version of this story is wrong.

Nobody failed to show up. The August 11 derecho put winds near 99 mph through Gary and knocked out power to more than 300,000 NIPSCO customers, the largest outage in the company’s history. Roughly 300 distribution poles and 44 transmission poles in Gary needed repair or replacement. More than 600 linemen were working in the city. Governor Braun declared a statewide disaster two days after the storm, and the president approved a federal Emergency Declaration covering 53 counties. The National Guard went to Gary, Munster, and Hobart.

That is a real response, and it arrived fast.

So this is not a piece about whether help came.

This is a piece about the order it came in. Gary is scheduled last. Full restoration by 11:59 p.m. on August 25, two weeks after the storm, later than every other community NIPSCO gave a date to.

Twenty-one years ago this country watched a predominately Black city drown on television and told itself the lesson was speed. Show up faster. Declare sooner. Do not let a president fly over at forty thousand feet.

By that measure, Indiana passed.

What nobody fixed was the ordering. And the ordering no longer has to defend itself as a judgment call, because it has been rewritten as arithmetic.

What actually got fixed

Let me give the country its due.

Hurricane Katrina made landfall on August 29, 2005. Thirteen months later, President Bush signed the Post-Katrina Emergency Management Reform Act, which reorganized FEMA and gave it new authority to close the gaps the storm exposed. FEMA got more autonomy. The FEMA administrator was elevated to serve as the president’s principal adviser on emergency management. New leadership requirements. New logistics systems. A national preparedness goal.

And it worked, at least on the thing it was built to fix. Five days from an Indiana derecho to federal paperwork. In 2005 it took nearly a week just to get water into a stadium.

Here is the part nobody put in a bill.

What actually decided who got their life back in New Orleans was not the declaration timeline. It was Road Home, the largest housing recovery program in American history. Road Home calculated your rebuilding grant as the lesser of two numbers: the pre-storm value of your house, or the cost to repair it.

That rule does not mention race. It does not mention neighborhood. It is a comparison operator. Take the smaller number.

But pre-storm home value in New Orleans was not a fact about houses. It was a record of eighty years of redlining, disinvestment, and appraisal practice. So homeowners in white neighborhoods, where pre-storm values ran higher, got grants closer to what rebuilding actually cost. Homeowners in Black neighborhoods got the lower number, even when the houses were the same size and age with similar damage.

One plaintiff in the eventual lawsuit received $1,400 to rebuild her home. Had the grant been based on repair cost, the Greater New Orleans Fair Housing Action Center calculated she would have received $150,000.

HUD settled in 2011 for as much as $62 million, covering roughly 1,300 Louisiana homeowners. Afterward, HUD stopped allowing that kind of loss-based payout to homeowners in future disasters. The formula was quietly retired, which is its own kind of admission.

A 2015 LSU study found that 70 percent of long-term white residents made it back to New Orleans within a year. For long-term Black residents, it was 42 percent.

Nobody had to say a single word about race for that to happen. That is the point. Every person administering Road Home could have passed a lie detector test about their intentions and the outcome would have been identical, because the discrimination was not in anyone’s heart. It was in the operator.

Take the smaller number.

The arithmetic

Which brings me to Melody Birmingham, an executive at NIPSCO’s parent company, and why her answer is not the reassurance it sounds like. Asked whether Gary was last because of race or income, she said no: “I will tell you that this is not a socio-economic divide.” She said she has family and friends in the city.

I believe her. She is telling the truth about intent.

She is answering the wrong question.

The right question is what the rule is. And NIPSCO has published it. Restoration priorities, the company says, are set by system needs, safety, and the work required to “restore service to the greatest number of customers as efficiently as possible.” In a briefing to another affected city, NIPSCO stated the order plainly: medical facilities and critical infrastructure first, then large customer counts, then commercial and residential.

Read it the way you would read code.

In practice, “restore the greatest number of customers as efficiently as possible” behaves like an optimization problem: it rewards work that returns more service for less time, material, and access complexity.

Gary has the largest number of customers in the state waiting. If the rule were simply most customers first, Gary would be first. But Gary also has 300 distribution poles and 44 transmission poles down, and Birmingham herself said the city sustained the “most extensive and complex damage” in the region. More customers waiting, and far more work required to reach each one. The second number governs.

The public explanation suggests a restoration logic in which damage complexity and crew productivity matter. When that logic meets a city shaped by long-term disinvestment, the result can be predictable: the residents facing the hardest repairs wait the longest.

Gary goes last. Not because Gary is Black. Because Gary is expensive.

And why is Gary expensive? Partly the storm, which hit where it hit. Partly something older. Gary has lost 61 percent of its population since 1960. The people left, and the poles did not. A shrinking city still has to hold up the same miles of wire, with fewer ratepayers to spread the cost of maintaining it across.

Let me mark where the record ends and I begin. What I can document is the population loss and the damage totals. I do not have data on the age of Gary’s distribution equipment, and no such data is published at the city level. Indiana reports reliability by utility, not by neighborhood. The line I am drawing between sixty years of leaving and a two-week wait is my inference, and you should hold it as one.

Here is what is on the record. In December 2024, in NIPSCO’s pending rate case before the Indiana Utility Regulatory Commission, an expert witness for the LaPorte County Board of Commissioners went through the IURC’s own reliability reports and found that from 2019 through 2023, setting aside major storms entirely, NIPSCO customers experienced longer average outages than the customers of any other investor-owned utility in Indiana. In the same filing: NIPSCO’s residential customers already pay the highest electric bills of any Indiana investor-owned utility, and the company was asking for a 20.1 percent revenue increase on top of the roughly 19 percent it was granted in 2023. She was an intervenor’s witness, which means she was paid to make the company look bad. Her source was the state’s own report.

So the two-week wait is not a freak event. It is the far end of a distribution that was already the worst in the state, at the highest price in the state, from a company asking for twenty percent more.

Then there is the second gear, and this one is worse.

Once the pole is back up, the utility’s job ends at your house. The weatherhead, that metal cap where the service wires meet the roofline, is your equipment, not theirs. If the storm tore it loose, NIPSCO cannot reconnect you. Under Indiana regulations and IURC-approved tariffs, you hire a licensed electrician, you get the repair certified, and then you call to be reconnected.

That rule is perfectly equal. Every household in Northwest Indiana faces exactly the same requirement.

It also functions as a means test. In a neighborhood where the lights came back on Tuesday, the houses still dark on Friday are the ones that could not produce an electrician and a check. A fixed-income grandmother on the east side does not have power because of the last ten feet of wire, and the last ten feet is billed to her.

Credit where it is due: NIPSCO is assembling a list of certified electricians, partly through the Indiana Electricians Union, to speed those repairs. It solves availability. It does not solve cost.

Here is what I want educators and administrators reading this to consider, because you are going to meet this exact structure in a vendor demo.

There is more than one way to write that objective function. You can minimize total customer-hours without power across the whole service territory, which sends crews to the cheap fixes first and mathematically guarantees a tail of people who wait two weeks. Or you can minimize the longest wait any single household endures, which sends crews to the worst-hit place first and slows everyone else down slightly.

Both are defensible. Both are optimization. Only one of them gets called efficiency, and it is never the one that starts in Gary.

The other thing that repeats

On August 15, two NIPSCO employees were scouting damage in Gary when someone fired at their vehicle. No one was injured. The company paused restoration work in the city, and crews returned later that afternoon with law enforcement support.

The shooting was real. Police have said it is unclear whether the crew was deliberately targeted, and NIPSCO itself has pointed to the storm damage rather than the shooting as the reason Gary’s timeline runs long.

Now watch what the story did anyway.

On September 1, 2005, the Associated Press moved a story headlined “Katrina Evacuation Halted Amid Gunfire.” It ran in at least 150 outlets around the world. Snipers on rooftops. Shots at rescue helicopters. Anarchy in the Superdome.

A month later, Knight Ridder reported that the Air Force, the Coast Guard, the Department of Homeland Security, and the Louisiana Air National Guard had not confirmed a single incident of gunfire at a helicopter. Governor Blanco later acknowledged the reported violence was “definitely out of proportion to the reality.”

The lesson people usually take from that is about misinformation. Fact-check harder. Trust official sources less.

I think the lesson is about reflex.

Because in Gary the incident is verified, and the institutional response was the same one: stop the help, add police, reclassify the emergency as a security situation. One vehicle. Twenty thousand households in the dark. The pause lasted an afternoon and it did not cause the two-week timeline. But it did something else. It handed the internet a reason to justify the delay, and the reason arrived faster than any crew, and it is the reason people in comment sections are still giving for why Gary waited.

A city becomes the explanation for its own abandonment. That is the part that repeats.

What ten days does

Shatiya Hardin has been cutting hair in Gary for twenty-five years. Her salon has been closed since the storm. She told the Associated Press that her neighborhood took the first few days in stride, and she explained why: “We’re always in survival mode in Gary.” She said the quiet did not last, and that a city should not have to live that way indefinitely.

I am not exaggerating about survival, and I am not writing about it from a distance.

My sister has muscular dystrophy. She is on a ventilator. She needs a reliable source of electricity to live.

Her backup generator was not intended for weeks of support. It was equipped to handle 200 hours of service before needing maintenance. Because of the conditions and the demand, it has not been easy for her to get parts. A network of people across the country has been trying to help.

The stress and fear of being helpless is crippling.

The first week did not produce an uproar because the people it happened to had already adjusted their expectations downward to accommodate it. That is not resilience. Resilience is what you call it from the outside. From the inside it is the absence of any reason to believe complaining works.

An optimization rule that puts the hardest-hit city last will keep running as long as the hardest-hit city has stopped expecting to go first. The math and the resignation hold each other up. Neither one requires a villain.

Three days ago the New York Times Magazine published a long piece by Christopher Cox about the transformers the entire grid depends on. They are hand-built, one at a time, to specifications that almost never repeat. The wait for a large one has gone from under a year to 128 weeks since 2021, with some orders running five years out. American factories make only about a fifth of what the country buys. Roughly three quarters of the distribution transformers already in service, the gray cylinders on the poles, are past the end of their expected life.

Which means there is a line.

And Cox reports that the hyperscalers have started paying to move up in it, so their transformers get built next. The load he describes is artificial intelligence. A frontier model in training pulls 500 megawatts or more. He names Google’s Gemini and Anthropic’s Opus.

I want to be careful about what I am claiming, because it would be easy to overstate. A substation transformer is not a wooden pole, and a wooden pole is not the metal cap on somebody’s roofline. Nobody’s data center delayed a crew in Gary. There is no line running from a server farm to my sister’s generator, and I am not drawing one.

What I am pointing at is the shape. The country is not short on the ability to move electricity. It is short on specific physical objects with long lead times, which means there is a queue, which means somebody decides the order. At the national level, that order is settled by who can pay to skip the line. In Gary, it is settled by a rule about efficiency. In a house on the east side, it is settled by whether you can afford an electrician for the last ten feet of wire.

Same operator. Three scales. The answer never comes out in Gary’s favor.

Katrina’s lesson was supposed to be that the country would respond to an emergency faster.

The country responded faster.

Nobody promised who would be responded to first. Or last.

Addendums

Added August 21, 2026, after publication.

A reader sent me the actual Commission order in NIPSCO’s rate case, Cause No. 46120. I had been working from the testimony filed in that case, not the order that closed it. Three things came out of reading the whole thing, and one of them is a correction I owe you.

One. What happened to the rate case.

I wrote that NIPSCO was asking for a 20.1 percent revenue increase. That is accurate as to the request. I did not tell you what NIPSCO actually got, and I should have.

The company filed in September 2024 seeking about $369 million, a 20.15 percent system-wide increase, on a requested return on equity of 10.6 percent. In February 2025, NIPSCO settled with the Office of Utility Consumer Counselor and a group of large industrial customers. The Commission approved that settlement without modification in March 2025.

The approved increase was $257,043,752, about $111.6 million less than the ask, roughly a 30 percent cut. Return on equity landed at 9.75 percent. For an average residential customer, the estimated bill increase is about 16.75 percent, phased in over multiple steps.

The settlement also held the residential monthly customer charge at $14.00 rather than raising it, eliminated the $50 deposit for customers receiving LIHEAP assistance, waived the $90 electric reconnection charge for customers cut off for nonpayment, delayed disconnections when temperatures are forecast below 20 or above 90 degrees, and created a bill assistance program funded by an annual $1.5 million shareholder contribution.

Only one party opposed the settlement, the Citizens Action Coalition. The Office of Utility Consumer Counselor, the statutory representative of Indiana ratepayers, signed it.

I stand by what the essay argued. But an ask without an outcome is a number chosen for effect, and I do not want to be the person who does that.

Two. Somebody already asked the question I ended on.

I closed the piece with a national transformer queue and a line about how the order gets settled by who can pay to skip it. I reached for a New York Times feature to make that point.

I did not need to. It is in the Gary rate case.

Several parties in Cause No. 46120 asked the Commission to open a separate proceeding on what the filings call “large or mega load customers.” Data centers. The Citizens Action Coalition’s witness argued that those customers should pay the full embedded cost of service, the same as everybody else, rather than only their direct incremental costs plus, in the settlement’s phrase, “some portion” of the existing system. He called the arrangement prima facie discriminatory treatment favoring large load customers, and said it warranted additional scrutiny.

NIPSCO’s witness responded that the concerns were hypothetical and premature, since the company had not yet signed a contract with any such customer.

The Commission agreed. It found a subdocket “neither necessary nor appropriate” and moved the question to a different proceeding.

So the settlement that governs your rates today commits NIPSCO to collect a large load customer’s direct incremental costs and some portion of the existing system. Neither of those terms is defined anywhere in the order.

The Commission held a public field hearing on this case in Gary on December 5, 2024. Eight months before the storm, in that city, the question of who pays what and in what order was raised, called premature, and set aside.

Three. There is already a list.

This one I am going to ask rather than assert, because the two things are not the same and I do not want to blur them.

In the same order, the Commission granted NIPSCO a waiver from a state rule that had required an employee to make an on-site visit before shutting off a customer’s power. With smart meters, NIPSCO can now disconnect remotely. The company expected roughly 205,000 of those meters installed by the end of last year.

The order carves out exceptions. Certain customers cannot be remotely disconnected: customers without a documented phone number or email address, customers who opted out of smart meters, and medical alert customers.

Medical alert customers. The company identifies households where someone’s life depends on the power staying on, and it protects them from the shutoff process.

That rule is about nonpayment. It is not about storms, and I am not going to pretend otherwise.

But the list exists. NIPSCO knows which addresses those are, and it treats them differently when it is the one choosing to cut the power.

So the question is whether that same knowledge shapes the order of restoration when the power goes out on its own. Not whether NIPSCO cares. Whether the information is used.

I do not know the answer. I know the list is real, because the state wrote it into an order, and I know that is a fair thing to ask a regulated monopoly out loud.

In my opinion, this is a step in the right direction, but it is also important to note that this is not a cheating detector. Anthropic’s own documentation on the limits of watermarking explains why.

How did this come about?

Anthropic signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content). Article 50 requires providers of covered generative AI systems to make synthetic output machine-readable and detectable as artificially generated or manipulated. The Code is a voluntary framework for demonstrating how providers may meet those obligations; it does not replace the legal obligation itself.

Be careful about the limitations:

A detected mark tells you content may have been processed by Claude. It does not tell you Claude wrote it. There are several ways that people use GenAI to interact with test - proofread, translate, summarize, and convert files, and the output carries a mark in all of those situations. A student who ran their own paragraph through Claude to fix the grammar produces marked text. So does a student who typed “write my essay” and pasted the result. The mark does not distinguish between them.

I often run my text through for a proofread - including this article. I can honestly say I wrote all but one paragraph this article. Claude cleaned up grammatical errors and wrote the paragraph about what other companies and platforms offer similar services. I proofed that paragraph for content errors. Pangram detects 0% human writing… which is both good and bad.

The absence of a mark tells you even less. Anthropic lists the ways generated content comes back clean: text from a model released before marking was supported, text that was heavily edited or paraphrased or translated, passages too short to carry a reliable signal, files whose metadata was stripped by a format conversion or a screenshot, and any product or file type where that marking type isn’t supported.

Google is the only other company watermarking text, through SynthID in the Gemini app, and its detector is still gated behind a waitlist for journalists rather than open to anyone teaching a class. OpenAI marks images and audio and offers a public checker for them, but ChatGPT text output carries nothing. Adobe, Microsoft, Meta, TikTok, and several camera manufacturers attach C2PA metadata to media files, which disappears on a screenshot or a re-save. Roughly 190 organizations signed the EU Code, 82 of them as providers, so agreement to mark content is not the constraint. Text is the hardest modality and almost nobody has solved it. A student who uses ChatGPT instead of Claude turns in prose that no tool currently claims to detect. The signal is asymmetric, and it fails in the direction that matters least to us. Clean text proves nothing. Marked text proves something small.

This distinction is particularly important because the EU AI Act itself recognizes that ordinary editing assistance is not the same thing as the generation of new synthetic content. Its transparency provisions include an exception for AI used solely for standard editing assistance that does not substantially alter the input or its meaning. (https://artificialintelligenceact.eu/transparency-rules-article-50/)

The signal is asymmetric, and it fails in the direction that matters least for academic-integrity decisions.

Clean text proves nothing. Marked text proves something small.

What this changes

Less than the announcement’s headlines may suggest. If a mark cannot separate a student who translated their own sentence from a student who generated an entire assignment, it cannot resolve an academic-integrity case. A watermark might justify a conversation, particularly if it conflicts with other evidence. It cannot establish authorship, intent, or a violation by itself.

The equity problem arrives quickly.

Students who use AI most often for legitimate and permitted assistance may be disproportionately represented among marked submissions: multilingual students, students with disabilities, students developing academic-English fluency, and students using tools in ways an instructor has explicitly allowed. Meanwhile, students who can substantially revise, paraphrase, or route output through another system may leave no reliable signal at all.

That is not a reliable basis for enforcement. It is a recipe for uneven scrutiny.

The practical response has not changed: design assignments for evidence of process rather than post hoc certainty about provenance.

- Ask students to submit notes, outlines, drafts, revision histories, or decision memos.

- Include brief oral, in-class, or recorded components in which students reconstruct their reasoning.

- Build disclosure into the assignment itself, so students have a legitimate way to describe what tool they used, for what purpose, and what they changed afterward.

- Treat discrepancies as invitations to ask questions—not as proof.

A disclosure statement can tell you more than a watermark: “I used Claude to translate my outline, then rewrote the English myself” is actionable pedagogical information. A detected technical signal is not.

Article 50 does not stop with AI providers.

Paragraph 4 also addresses deployers: organizations that publish AI-generated or manipulated text to inform the public on matters of public interest generally must disclose that fact. There is an important exception where the content has undergone genuine human review or editorial control and a natural or legal person takes editorial responsibility for publication.

For universities, this is primarily a communications and governance question, not a classroom-detection mandate.

Institutions publishing public-facing, AI-assisted content should clarify who conducts editorial review, who assumes responsibility for the final material, and when disclosure is appropriate. Universities with substantial EU operations or public-facing EU communications should review those practices with institutional communications staff and legal counsel.

Provenance is not Authorship.

Origin marking is a context signal. Like image metadata, it can offer useful information about a piece of content’s history. It is reasonable for the industry to build better tools for that purpose.

A detected Claude watermark may support a conversation with a student. It should not, by itself, trigger an allegation, grade penalty, or disciplinary finding. Academic-integrity policies can acknowledge watermarking as one limited source of contextual information, but they should state plainly that it is not dispositive evidence of authorship or misconduct.

The goal is not to become better at catching marks. It is to build assignments, policies, and classroom relationships that make students’ thinking visible.

I have a lot to share about what I’ve been learning when it comes to AI, and just as much about what I still want to learn. That’s coming.

In the meantime, I want to share a chapter I had the privilege of working on for Seeing Interaction Design, the edited volume by Shunying An Blevis, Eli Blevis, Makayla Lewis, and Miriam Sturdee, out this month from Springer. My chapter is called “Examining AI-Generated Art in the Educational Context,” and it brings together three parts of my life: arts and craft (I’m a quiller), generative AI, and my time in Sydney as a Fulbrighter (which I work into as many conversations as possible).

This chapter provides a brief survey of the academic literature surrounding some of the essential questions related to AI generated art for art and craft educators. The author also examines practical advantages and challenges of using generative AI tools for design as an educator as well as a professional crafter.

I’d love to hear what you think.

A math professor stood in front of his class, finished proving that the square root of two is irrational, and nearly hugged himself.

He hadn’t planned it. The proof is old, older than almost anything else we teach, one of those arguments the ancient world handed down intact, and he had shown it a hundred times. But something about it caught him that day. He almost threw his notes in the air and wrapped his arms around his own shoulders like a man whose team had just scored. Most of the room, he admitted, had not followed every step. Didn’t matter. What they caught was the joy.

His name is Amitabh Chaudhary, he teaches computer science at the University of Chicago, and I was in the room when he told that story. It was the inaugural CRA-E Chicago-Area Regional Summit for Teaching-track Faculty in Computing, held in Chicago on June 30, 2026, and coordinated by Borja Sotomayor, Logan Paul, Jennifer Campbell, Jeff Turkstra and Kayley McDonald of CRA. The session was a series of lightning talks, a handful of instructors taking the stage in turn to name the thing that makes teaching stick. Chaudhary has a name for it. He calls it X.

Here is how he set it up, and it is an acute observation about AI in education that is easy to overlook. The job has two halves. One is efficient teaching: preparation, clarity, organization, being on time, having the humility to say “great question, I don’t know” instead of bluffing. AI is useful for that half. The other half is X. And AI, in his words, only pretends to be useful for X.

If you have watched a chatbot generate a clean, well-sequenced lecture on a subject you spent years learning to teach, you have probably felt the small cold thought that comes next: what is left for me? Chaudhary’s answer is that the machine took the half of your job you were allowed to be tired of. What remains is the half you got into this for. That is not a consolation prize. On a talent show, the X factor is the quality the judges can’t define but recognize the second it walks onstage. Teaching has the same thing. We just spent a long time pretending it wasn’t real, because you can’t put it on a rubric.

Let me tell you what X actually looked like from my seat that afternoon, because “be passionate” is useless advice. I will try to explain it using evidence-based practices, and provide a list of resources at the end of this Substack.

It looked like realness you can’t rehearse.

The hug was one version. Another came from Chaudhary’s algorithms class. He had time left at the end of a session, opened the textbook to a problem he had never seen, and tried to solve it live in front of everyone. He did not get far. He said this like it was the whole point, because it was. Students learn something real watching you walk into a wall and stay there and try the next thing. They are not learning the algorithm in that moment. They are learning that a competent adult can be stuck and survive it.

A language model does not get stuck. Ask it to, and it will perform a very convincing version of stuck, complete with “hmm, let me reconsider.” But it was never actually lost, and some part of a student clocks the difference between a person working without a net and a system running a script about working without a net. The realness is the entire transaction. You cannot fake it, which is exactly why it lands.

It looked like a bet on the specific people in the room.

For a few years, Chaudhary was afraid to teach the dual formulation of support vector machines. It is hard. He worried it was too hard, so he left it out, and his students graduated believing it was beyond them, because leaving it out had quietly told them so. Then he stopped flinching and taught it. Students took it in stride. Now they finish his course assuming that level of difficulty is just the water they swim in.

This is the move AI is worst at, and it is worth being precise about why. A model is built to return the most probable, most broadly acceptable answer. It regresses to the mean by design. It will cheerfully meet a class wherever it guesses the average student is standing. Raising the bar is the opposite reflex. It is a wager on particular humans that they can clear more than the average predicts, and it only pays out because a person who knows them placed the bet out loud. Zaretta Hammond, in Culturally Responsive Teaching and the Brain, calls this the warm demander: high expectations carried on a real relationship, neither half working without the other. Relationships, in her phrase, are the delivery mechanism for rigor.

And it looked like the tell.

In the same session, Dale Reed from the University of Illinois Chicago (UIC) walked us through an exercise from a course he co-teaches: read three-sentence stories, guess which came from a person and which from a chatbot. His students had gotten good at it, and they had found the giveaway by measuring instead of vibing. The single most reliable signal was average sentence length. The AI’s sentences came in around twelve words, tidy and even, every single time. It writes to the middle. That evenness is the fingerprint.

Which tells you everything about X, because X is the opposite of the middle. It is the twelve-word sentence sitting next to the forty-word one. It is the professor who hugs himself, the one who gets stuck, the one who decides you are ready for the hard thing. High variance, specific, embodied, sometimes a mess. The machine’s smoothness and the teacher’s X are not two points on one scale. They point in opposite directions.

Baker Franke, also from UIC, had students build little decision trees by hand to sort Chicago neighborhoods, north side or south side, then ran mystery data through their handmade models and revealed, at the last second, that the data was actually South Korea. The room came apart. Some students leapt up. Some slumped like they had just missed a penalty kick in the World Cup. You do not get that from a chatbot. Not because it can’t run the numbers (it can, faster) but because nobody leaps out of a chair over a thing that was only ever efficient.

Chaudhary ended on his least polished note. His students had complained that his assignments were too long. It stung, and he walked into class still stinging. Then he told them the truth: the assignments are long, the complaint is fair, and if you want to keep up in this field you have to learn a great deal very fast, so curse me if you need to, but do the work. He had retitled his own talk that afternoon, too, after watching the teacher who went before him. X, he said, you could just call Baker.

So here is the one thing to try. When the semester starts, teach one thing live and unrehearsed. Pick a problem you have not pre-solved, put it on the board, and work it in front of them. Let them watch you not know, and then watch you stay. It is a real demonstration of persistence, not a performance of one.

Aside: Leon Johnson, from Indiana University Indianapolis, gave a great talk in the same session on specs grading. I’m not including it here because he agreed to present it to Luddy faculty in the upcoming academic year, and I’ll have a post dedicated to his presentation at that time.

A note on the evidence

The firmest support is for two of these. Emotion is what sticks: excitement around the moment of learning strengthens what makes it into long-term memory, including a study where arousal after a lecture raised exam scores. And students learn from struggle, their own and an expert’s, the through-line of the productive-failure and vicarious-failure research. Two deserve a lighter hand. A teacher’s visible enthusiasm reliably raises motivation and enjoyment, though the link to grades is thin (Keller et al., 2018; and here). High expectations do pull students up, but the effect is a contested nudge rather than a lever (teacher-expectancy effects on achievement).

A joy-centered reading list

A few books and resources on joy-, care-, and emotion-centered teaching, if any of this struck a chord.

Essential reads for joy-centered teaching

• Joy-Centered Pedagogy in Higher Education: Uplifting Teaching and Learning for All, edited by Lori Gonzalez (open access). A collection whose reflection questions and teaching tips turn joy from theory into concrete, book-club-ready practice.

• A Pedagogy of Kindness, Catherine Denial. A repeatedly recommended read that challenges our assumptions about “rigor” and reframes what humane, meaningful teaching can be.

Institutional transformation and care

• Hope Circuits: Re-wiring Universities and Other Organizations for Human Flourishing, Bessette. On doing hopeful, justice-oriented work inside fraying institutions, and rebuilding those institutions for human flourishing.

• The Caring University, Kevin McClure. A research-informed case that care (real practices, structures, and leadership habits) is fundamental to educational excellence, not soft or secondary.

The science and practice of emotional engagement

• The Spark of Learning: Energizing the College Classroom with the Science of Emotion, Sarah Rose Cavanagh. A foundational look at the science of emotion in learning, where a teacher’s contagious passion becomes the best route to engagement.

• The Joyful Online Teacher: Finding Our Fizz in Asynchronous Classes, Flower Darby (forthcoming April 2026). Brings joy-centered pedagogy into asynchronous online courses, with tips that adapt to in-person and synchronous classes too.

• The Present Professor: Teaching, Engaging, and Thriving in Higher Ed, Elizabeth A. Norell. Argues that building psychologically safe, inclusive spaces for students begins with educators sorting out their own identities first.

• The work of Peter Felten and Mays Imad. Anything by Felten (relationship-rich education, high-impact practices) or Imad (trauma-informed, hope-centered pedagogy) for holding rigor, humanity, and both distress and joy in the classroom.

Practical resources and creative inspirations

• Dartmouth JoyCards (free PDF). A “sweet little” set of fun, practical joy-filled teaching techniques from Dartmouth’s Design Initiative, perfect supplementary material for book clubs.

• The Art of Gathering and Unreasonable Hospitality. Picks outside the usual teaching-and-learning fare that offer fresh perspectives on creating meaningful, joy-filled shared experiences.

but instead, taking advantage of some of the resources Eleven Labs provides for educational means.

Thirty minutes later I had Aria. Warm voice, female, American accent, fluent in twenty-five-words-or-fewer customer service for a fictional dentist named Northwind. The Voice Agent Academy I developed had me pick her from a panel of six voices. I named the agent, wrote her a system prompt, gave her a knowledge base, and watched her come online inside my browser.

The strange part was what happened next.

You are a warm, efficient front-desk agent for Northwind Dental.

I started remixing. You are a warm, efficient front desk for my course. Two beats later I was picturing the syllabus, the late policy, the specific student last semester who emailed me at midnight to ask whether the midterm was cumulative. (It was. We had discussed this. The syllabus was bolded.)

The build took thirty minutes.

The rest of this piece is about who can actually reach her.

What TikTok already taught us about UDL

We just scoped what the agent does. The harder question is who it reaches, and pop culture has worked out part of that answer already: scroll TikTok, scroll Reels, and captioned video is everywhere, one tap to enable, so thoroughly normalized that most of us stopped noticing the shift happened at all. Captions stopped being an accommodation and became a baseline, one that serves the deaf and hard-of-hearing community, language learners, autistic viewers, parents watching at 2 a.m. with a sleeping baby on their chest, and most of us most of the time. The analogy has limits, since entertainment captures attention while courses ask for sustained cognitive work, but the access principle holds, and it is the question voice agents in education will be judged by: was the option for students who cannot use voice intentional, or an afterthought built from complaints?

The Captioning Standard

Out of the captioned-by-default observation comes a test worth naming. Call it the Captioning Standard:

If this voice-based learning experience were converted into a captioned, text-first format, would it still deliver equivalent cognitive value?

If yes, the voice version is an enhancement, and a captioned transcript is your parallel build. If no, voice is doing irreplaceable cognitive work, which means you owe students who cannot access voice an equivalent learning experience, not a transcript pretending to be one.

Try it on the oral practice partner use case. A nursing student rehearsing a patient hand-off cannot get equivalent cognitive value from a written transcript of someone else’s exchange. The skill is the speaking, the timing, the under-pressure phrasing. The standard tells you, before you build, that you need an equivalent rehearsal modality (a partnered human simulation, for instance) for students who cannot speak out loud, not a transcript posing as access.

The five use cases the Academy mapped

The Academy walked me through five educational deployments. A syllabus hotline for logistics questions. A voice tutor that lets students work problems out loud. An oral practice partner for OSCEs, moot court, and language exams. A five-minute retrieval check-in where students get quizzed by phone after a chapter. And what one of the tutorial screens called, with admirable honesty, the after-hours line: the 2 a.m. call when a student is stuck and no human is awake.

Each one is a realistic scenario that can serve a different kind of student differently. Once you see them through the captioned-TikTok lens, the question stops being whether to deploy and starts being where in the matrix each one belongs.

A Pedagogical Fit Matrix

Here is the mistake most of us will make: we will start by building. Start by placing.

UDL gives you the access baseline. Differentiation tells you which tool to use, when, and with whom. The matrix below is differentiation in Tomlinson’s sense, applied to course design rather than to individual lesson planning. Different use cases serve different students doing different kinds of cognitive work. Your job is to choose among them deliberately.

Two axes. The first is what the agent is asking the student to do cognitively. Logistics, retrieval, rehearsal, simulation. The second is how many students can reach the learning through voice without a parallel build.

Syllabus hotline Cognitive function: Logistics Access alignment: Wide Where to start: First pilot

After-hours line Cognitive function: Mixed Access alignment: Wide Where to start: Early pilot

Retrieval check-in Cognitive function: Recall production Access alignment: Medium Where to start: Pair with text recall

Voice tutor Cognitive function: Rehearsal Access alignment: Medium Where to start: Pair with chat tutor

Oral practice partner Cognitive function: Simulation Access alignment: Narrow Where to start: Design the parallel before the build

Choose one. Don’t overwhelm yourself.

To use it: pick the use case you are considering, read the cognitive function (what learning work the agent is doing), check the access alignment (how much parallel build you owe students before deployment), and start where it tells you to start.

Start in the upper rows. Move down only after you have built the parallel for students the voice option excludes. The oral practice partner is the most pedagogically interesting use case for nursing, language, and law programs, and also the one that demands the most careful equity design, because the production of voice is the skill being assessed.

This is a feature, not a bug. Different disciplines need different positions because they ask students to do different kinds of cognitive work. A studio professor does not need what a quantitative methods professor needs. The question is never which voice agent we adopt at the institutional level. The question is which use case each course needs, and what pedagogical support faculty need to figure that out. Differentiation, taken seriously.

When voice is the wrong tool

I will tell you, plainly, what I would not use voice agents for.

Anything summative. FERPA exists for a reason, and the minute your agent is anywhere near assessment or accommodation data, you need institutional review, not a Saturday afternoon. Voice agents are for low-stakes formative practice. They are not for high-stakes assessment.

Anything where confident accuracy matters and the consequences of a hallucinated answer are real. Voice agents are confident when they are wrong. I have spent enough time around the build process now to know that it’s easy to miss nuances that can cause an agent to invent a deadline, misstate a chapter, or fabricate a policy. I am working through more formal programming training partly for this reason. I am fortunate to work in a school of computing where colleagues will patiently help debug. Not everyone has that safety net, and I do not want to pretend they do.

Anything that involves giving an agent access to your folders without you fully understanding what is in them. Pointing an agent at a directory hands it FERPA files, IRB materials, draft manuscripts, tax documents. Even with training opt-outs, contents travel to vendor servers. Most universities have not issued guidance that keeps pace. There are some circumstances where faculty are navigating this alone, and “I built it myself” is not the same as “I vetted it.”

And there is a pedagogical question I have to keep in mind. When an agent generates the question a student should have asked, who is doing the cognitive work? The line between scaffolding and substitution is thinner than the marketing copy suggests. I do not want to teach students the theater of inquiry. I want to teach them the practice. “The student decides whether to act on the feedback” can quietly become the alibi that lets us skip asking whether the student actually noticing when they should inquire is the skill we wanted to build.

What to build alongside

If you do pilot, plan the parallel from the start. Pair voice with a captioned transcript so deaf and hard-of-hearing students get equivalent content. Pair it with an asynchronous text channel for students who think better in writing or cannot speak out loud in shared housing. Pair it with home-language support where your student body warrants it. Yes, this is extra work, but not as much as it used to be with voice to text software and other tools… but I don’t want to minimize that for instructors with large teaching load. I will discuss that a bit more in the next section.

Voice recognition is still not equally good for everyone. It performs worse on speakers of African American English, on non-native accents, and on speech with regional variation. If voice is the only door, the door does not open for some of our students.

Sometimes the inclusive tool is a well-moderated discussion board and a printed one-page FAQ. The boring tools earn their place by working for everyone.

If the parallels feel like more work

If you are teaching a section that grew from 30 to 120 students with the same support staff and the same office hour budget, the parallels above can read like a fourth job stacked on top of three you already cannot finish. That overwhelm is real and mostly not your fault. Larger sections and shrinking resources are an institutional decision, not a personal failure to keep up.

A few honest things about the work.

Most parallels are not new content. They are labeling and routing what you already have. Your syllabus is the text FAQ. Your LMS discussion board is the asynchronous text channel. Multilingual support is the genuine exception, since most institutions do not actually provide translation services. If your student body needs that layer, it is a real cost worth surfacing to your chair or your provost rather than absorbing alone.

The matrix keeps the work proportional. The wide-access use cases have parallels that take minutes. The narrow-access use cases take a semester to design well, which is why the matrix recommends not starting with those until you have the runway.

And the harder version of the answer: if you do not have time to build the parallels, you do not have time to deploy the agent. A voice-only pilot quietly shifts the cost of your overloaded course onto the students the agent leaves out. Their experience becomes the silent line item that absorbs the scarcity. Better to wait a semester and build the parallels right than to ship something that creates a new inequity inside a course already strained by the old ones.

Where I want to take this

If you are a faculty member or instructional designer thinking about where voice agents belong in your course, I would like to keep working on this with you. Reply or subscribe at akesha.substack.com to follow along.

A first step for this week

Pick the use case in the wide-access top row of the matrix that touches your inbox most. For most of us, the syllabus hotline might be a better alternative than constantly repeating, “It’s in the syllabus.” Build a small agent trained on your syllabus, your late policy, and your assignment FAQ. Work with your IT team to pilot a study so you don’t run in violation of FERPA and other policies. Tell students it exists, tell them it can be wrong, and offer the same content in a written FAQ for anyone who prefers to read.

Run it for four weeks. Log three things each week: error types (what categories of mistakes the agent makes; make it a game, since students love correcting you, and perhaps the student who identifies the most mistakes can win a prize), usage patterns (which students use it, when, for what), and unmet needs (questions students bring that the agent cannot answer well).

What you will learn in four weeks is not whether voice agents work. You will learn which version of differentiation your students actually needed.

The strongest deployment of AI agents in a course I have heard described is not one where students use them. It is one where students design them, place them in their own version of this matrix, and contest whether the placement is right. That treats AI as something to be made legible and contestable, not consumed.

That is what I want for us as a field. Legible, contestable, parallel-built. And available, deliberately, to the students it serves best, with real plans for the ones it does not.

The voice agent question isn’t can. It’s for whom. And the answer is never everyone.

I know how that sounds. The gap between last year and tonight is the point. The interesting thing about agents isn’t inbox triage. It’s that they can take over the routing that quietly killed some of our best teaching. That routing was always too much for one person to keep doing by hand. I’m talking about classrooms here, but the pattern shows up anywhere your job is half thinking and half shuffling work between people.

For about a year, most of us have used AI the way we use a search bar. You ask, it answers, you copy the part you wanted. Useful. Also forgettable. What’s shifting now is the move from talking to a model to handing one a job and walking away. A chatbot is a resource you consult. An agent is closer to a coworker you delegate to. While this isn’t the first time I have discussed AI workflows:

That difference matters a lot more in a classroom than the productivity crowd lets on, and I want to explain why.

First, the small win everybody’s selling

You’ve seen the pitch. Connect an agent to your email and calendar, have it scan your inbox every morning, and get back the five things that actually need you, your schedule, and where your open hours are. Run it again in the afternoon. People put a dollar figure on it: save an hour a day, value that hour at thirty bucks, and you’ve reclaimed something north of ten thousand dollars a year. An hour a day across a working year is real time.

Handing an agent your inbox also means it can read everything in there and, if you let it, act as you. That is its own risk, and I’ll come back to it.

I use a version of this. It’s genuinely handy, and if your mornings disappear into triage, set one up. But here’s my honest read: the inbox briefing is the least interesting thing an agent can do. It’s a faster version of something you already knew how to do. It doesn’t change what’s possible. It just changes how long it takes.

The thing that changes what’s possible is when an agent takes over work that was logistically impossible for one person to sustain. That’s the part worth your attention, especially if you teach.

What an agent actually is

Here’s what I mean by “agent.”

A plain language model is a cookbook. It holds every recipe and every technique, and it’ll answer any question you ask, but it just sits on the shelf. An agent is the line cook. You don’t hand a line cook a script, you say “dinner for eight, plated by seven,” and they read the orders, prep, cook the dishes in order, taste as they go, and fix what’s off. Three things make that possible. It works from the goal you set instead of waiting for the next instruction. It remembers how the kitchen runs, so it isn’t relearning the menu every shift; it holds onto your preferences and what happened last time. And it has hands. A cookbook can tell you how to hold the knife; the line cook actually picks it up. That’s the difference that counts: an agent doesn’t just tell you to send the email, it sends it, updates the file, pulls up the calendar.

That’s the core loop. Reason, act, check the result, try again. Less search bar, more coworker who learns the job.

Revisiting peer responses

Here’s where this stops being a productivity tip and starts being about teaching.

Peer review is one of the best things you can do in a technical course. Students read each other’s code, run it, pick at the logic, and learn to see their own work by digging into someone else’s. In a lab together it works because the logistics are invisible. Everyone’s there. The code is up on the screen. The feedback travels because the people are three feet apart, leaning over the same laptop.

Move it online and asynchronous and it usually dies. Not because the pedagogy stops working. Because the routing turns into a second job. The teaching is the cooking: reading the code, running it, explaining what breaks and why. The routing is all the prep and plating around it, collecting each submission, matching it to the right reviewer, sending it on, chasing the three students who haven’t turned theirs in, gathering the reviews, bundling them, and getting the bundle back to the author before the moment passes. With no line cook to hand that off to, the chef does every bit of it alone and never gets near the stove. That’s the online instructor. Across a class, the whole thing collapses under the coordination. So we quietly drop peer review and assign something easier to run, and students lose the part that taught them the most.

This is exactly the kind of workflow an agent can carry. The clearest worked-out version I’ve seen comes from a writing classroom rather than a technical one, but the pattern is the same; only the artifact changes. The educator Michelle Kassorla has mapped out a feedback loop for her asynchronous writing courses, a design she calls the “Full-Loop” Routing Feed, built to run inside Google Workspace for Education. She hasn’t deployed it yet; she’s waiting on her institution’s access. But it’s the first time I’ve seen an agent not just save time but bring back a high-value teaching practice we’d quietly given up on. She gives each step in the hand-off a named role. The Vault holds the writer’s opening materials, their recorded read-aloud of the draft and a short “face-saving” comment about what they were going for. The Aggregator bundles those clips with the main respondent’s marginalia and critique into one review package. The DJ drops that package into the right group members’ to-do feeds. The Router gathers everyone’s recorded reactions. The Closer returns the finished bundle to the writer’s private inbox.

It’s the same sorting and shuttling I used to do by hand, except now it runs itself, on a schedule, without dropping anyone. Swap the draft for a pull request and the read-aloud for a recorded walkthrough of the code, and it’s the same loop in my course. The students never see the sorting. They see “here’s the code you’re reviewing” and “here’s what your group said about yours.” What they’re left to do is the only part that needed a human in the first place: read the work closely, respond honestly, and revise. The agent took the file management. The students kept the thinking.

Not “AI does the teaching.” AI does the logistics that were taking away time from the teaching.

The part the productivity crowd skips

Here’s the caveat I promised back at the inbox agent. Giving one access to your own email is a risk you take on yourself. Giving one access to a classroom raises the stakes, because now it’s not just your data on the line. An agent is a capable intern, not a colleague you stop checking on. Three things go wrong, and they go wrong in ways that land on students.

Agents drift. On a long task an agent can lose the thread of what it was asked to do. Tell one to draft a note about a project running late and, after enough self-editing, it can hand you a cheerful update saying things are going great. Now picture that drift inside a feedback loop. Whose draft gets routed to the wrong group? Whose comments arrive attached to the wrong name?

Agents are confidently wrong. They’ll do the thing and report success with total certainty, even when the output is invented. A wrong answer delivered with a straight face is harder to catch than one that looks unsure, and students tend to trust the confident voice.

And agents touch private things. This is the one that really depends on what your school’s policies are on whether and how AI can interact with student work. It’s your responsibility to know, and to make sure your students understand how AI interacts with their work, and whether they’re okay with it. Can they opt out if they’re not? That is a different conversation, but a necessary one. To run that feedback loop, the agent has access to student work, student writing, sometimes student recordings. And student work is not just sensitive, it’s regulated. Under FERPA, drafts and the feedback attached to them count as education records. So where you send them isn’t just a tech choice, it’s a legal one. An agent running inside your institution’s approved Google Workspace, under a data agreement the school already vetted, is one thing. Someone bolting a personal AI account onto a stack of student drafts is another, and it’s the kind of shortcut that looks efficient right up until it isn’t. A sloppy permission or a careless prompt and the agent shares something it shouldn’t, or edits the wrong file. When the data is a 19-year-old’s first vulnerable draft, “oops” is not a recovery plan.

There’s a partial way out of that last problem, and it’s worth knowing about. In a recent piece in Medical Teacher, Ken Masters and colleagues make the case for small, locally-hosted language models in education. The idea is plain. Instead of shipping student work to a company’s cloud, where it can become training data, leak, or be recombined to re-identify someone, you run a smaller open-source model directly on your own machine. Nothing uploads. Running it locally puts you on roughly the same footing as a statistics package installed on your laptop, which is a far easier conversation to have with FERPA, HIPAA, or your institution’s risk office. The models are small enough to run on a well-specced computer without a server farm, and they keep working with the wifi off.

Here’s the trade-off. Smaller models are slower and less accurate, and they get shaky when a task needs open-ended reasoning across different domains. Masters and his coauthors found something that will sound familiar: in thematic analysis the models named the right themes but invented the supporting quotes, catching the sense of what a student said while fabricating the actual words. The confident-wrongness problem again, just on your own hardware this time. Read the article if any part of your workflow touches student data; it’s the most practical thing I’ve seen on the privacy question.

So the human stays in the loop, but the job changes. You’re back to being the chef: you set the menu, decide what the agent is allowed to touch, and taste before anything leaves the kitchen. Judgment doesn’t get automated. It gets concentrated into the decisions that actually need it.

From using AI to managing it

A year ago I was asking AI questions. Now I’m delegating jobs to it and checking its work. It’s a different relationship and a different skill. When I treat AI like a search bar, I ask a question, copy the answer, and I’m still doing all the routing and judgment myself. When I treat it like a capable intern, I hand over a goal and a pile of logistics, then stand at the end and taste the work before it reaches anyone. The thing to get good at isn’t prompting. It’s managing: knowing what to hand off, what to keep, and where to stand guard.

If you want to try it once, pick a single task in your teaching that dies online because of routing: peer feedback, matching project partners, collecting and returning work on a schedule. If you don’t teach, swap peer feedback for whatever task in your own work dies in email and spreadsheets. Write down only the parts that are pure logistics, the collect, match, remind, and return. Hand those to an agent, even a clumsy one stitched together from a shared folder and a simple script, and keep for yourself the parts that need judgment. Run it for one assignment, then sit down with your students and look at what broke. That review is the actual lesson, for you and for them.

So the question I’d leave you with isn’t whether the technology works. It does, unevenly, with supervision. The question is which tedious, logistics-heavy thing you’d hand to an agent so you could give your students, or yourself, the part that was always the point.

For me it was the routing. What’s yours?

Most conversations about generative AI in programming courses still run on instinct. One camp wants to ban the tools to protect the fundamentals. The other wants to adopt them quickly so students are ready for an industry that has already changed. Both positions get argued from conviction, and both skip the step that should come first: looking at what the research says.

That step is now possible. In the past two years, computing education researchers have produced enough studies to move the question from opinion to evidence. The findings do not hand a clean win to either camp. They support bringing AI into the programming classroom, and they show that doing it badly can widen the gap between students who are already succeeding and those who are struggling. The case for integration and the case for careful design turn out to be the same case.

What the tools are good at

Start with the capability question, because for the types of problems we usually assign, it is no longer in serious dispute. Current large language models solve the kinds of problems we assign in introductory courses. A review of the field by Paul Denny and colleagues documents how code-generation tools went from a curiosity to a standard part of the programming workflow in a matter of months, and how readily they handle the small, well-specified problems that fill most CS1 assignments. The exam evidence is concrete. When James Finnie-Ansley and colleagues ran OpenAI’s Codex through two CS1 tests in 2022, its answers ranked seventeenth in a class of seventy-one, inside the top quartile, and that was an early model. Newer ones have only widened the margin.

The industry signal points the same direction. In April 2025, Microsoft’s CEO said as much as 30% of the company’s code was being written by AI. That figure is a remark on a conference stage, not a measured result, and “written by AI” has no agreed definition, so it deserves caution. As a marker of where professional practice is heading, though, it is hard to ignore. If we want graduates who can work the way their future colleagues already work, ignoring the tools is its own kind of risk.

The second strength is tutoring. A first-year case study by Liam Stienstra and colleagues used a ChatGPT-based tutor to teach a new topic, then tested students without assistance. The tool supported learning, and most students preferred the AI tutor to a human one. Their reasons are worth considering: it was always available, and it did not make them feel judged for asking a basic question. Purpose-built tutors push this further. Tools like CodeHelp, from Mark Liffiton and colleagues, and CodeAid, from Majeed Kazemitabaar and colleagues, wrap the model in guardrails so it nudges a student toward an answer instead of handing one over. Work by Harsh Kumar and colleagues, run across a classroom study and a larger controlled experiment, found that the structure an instructor puts around the tool changes how students use it. Guidance that prompted students to attempt a problem before turning to the model, or to reflect on their own thinking, cut down on aimless queries and on the habit of pasting an assignment straight into the chatbot verbatim without attempting it first.

There is also early evidence that the help does not vanish the moment the tool is taken away. In a controlled study, Kazemitabaar and colleagues gave sixty-nine beginners, ages ten to seventeen, a sequence of forty-five Python tasks. Half had an AI code generator during the authoring work. That group completed more tasks and scored higher on the authoring work. The natural worry is that they would fall apart once the tool was gone. They did not. On the modification tasks they had to do by hand, they performed no worse than the students who never had access. On a retention test a week later they came out slightly ahead, though the difference was not statistically significant. Encouraging and modest at once, which is the honest way to hold it.

So far this is the authentic optimistic story. It is also incomplete.

The catch is an equity problem

Learning with AI is not automatic, and the students it fails are the ones we should worry about most.

The clearest evidence comes from James Prather and colleagues, in a study with the apt title “The Widening Gap.” Using eye-tracking and think-aloud methods with novice programmers, they found that every metacognitive difficulty students had before AI is still present, and that new ones have appeared. The pattern that is concerning: students with higher grades and stronger self-efficacy used the tools to accelerate, while weaker students were often slowed down and left with what the authors call an illusion of competence. The work finished but the understanding did not. The same tool that helps a confident student move faster can help a struggling student fail without noticing.

Part of the reason can be found in a paper by Lev Tankelevitch and colleagues on the metacognitive demands of generative AI. Working with these systems asks the user to do things that are cognitively hard: state a goal precisely, break it into parts, judge whether the output is any good, and adjust course when it is not. The authors compare it to a manager delegating to a team. That is a demanding skill, and it is not one novices arrive with. Studies of how beginners actually use these tools back this up. A close look by James Prather and colleagues at novices working with Copilot found students leaving out details that matter, like input and output types, and, when a prompt failed, changing a few words in the prompt rather than supplying the missing requirements.

This is where the equity lens is needed. A tool that rewards students who already plan well, monitor their own understanding, and know what they do not know will tend to reward the students who were going to do fine anyway. Left unaddressed AI creates a larger equity gap.

What evidence-based design looks like

The same body of research points to what helps, and the through-line is consistent: slow students down enough to make them think, and build the supports that strong students supply for themselves.

Majeed Kazemitabaar and colleagues tested this directly. They compared a baseline that simply showed students AI-generated code with an explanation against seven techniques that required some engagement with the code before or after it appeared. The technique that worked best, called Lead-and-Reveal, brought what students thought they had learned closest to what they actually learned, without adding cognitive load. The effects were modest and the samples small, which the authors are candid about, but the direction matches everything else in the literature. Passive acceptance of generated code produces the illusion of learning. Forcing a small act of reasoning interrupts it.

Other design moves follow from the same logic. Scaffold the tool rather than handing it over, as Kumar’s guidance strategies did. Teach prompting as an explicit skill, since novices do not pick it up on their own. Move assessment toward process and away from the take-home artifact that a model can now generate in seconds. And give upper-level students structured practice with the messier task they will actually face at work: a study by Anshul Shah, Leo Porter, and colleagues taught students to use GitHub Copilot inside a large codebase and found a pattern they call one-shot prompting, where students ask the tool to build an entire feature at once and then spend their effort debugging what comes back. That is a teachable habit, and naming it is the first step to correcting it.

There is a quieter finding that reframes the whole debate. A 2025 replication study by Rita Garcia and Michelle Craig revisited a 2004 survey of what CS1 instructors teach and find hard to teach. Twenty years on, the basic concepts are largely unchanged, recursion is still the hardest thing to teach, and the newest challenge instructors name is teaching students to think and plan before they code. That difficulty predates generative AI. The shift toward planning, decomposition, and self-monitoring, the very skills AI makes more important, was already underway. The course it was happening in was already struggling: introductory programming has carried high failure rates for decades, documented by Jens Bennedsen and Michael Caspersen and revisited by Christopher Watson and Frederick Li, and the long line of work on Soloway’s rainfall problem, surveyed by Kathi Fisler, showed that many students finish CS1 unable to write a basic loop correctly. AI did not break the introductory course. It exposed problems the course already had.

You do not have to figure this out alone

The research is scattered across conferences and journals, which makes it hard for any individual instructor to assemble. That gap is part of what the Consortium for Generative AI in CS Education https://www.teachcswithai.org/about exists to close. Housed at UC San Diego’s Center for Research on Education, Assessment, and Teaching Excellence and led by Leo Porter, Daniel Zingaro, and Beth Simon, the consortium gathers course materials, research summaries, and a community of educators working through the same questions. Its partner network runs from professional bodies like ACM and SIGCSE to organizations focused on access and inclusion, which matters given what the research shows about who gets left behind. The point of the consortium is not to sell a position. It is to make the evidence usable, so that a faculty member redesigning a course in the summer is not starting from a blank page. About half of the articles I used for this article can be found in the consortium’s resources section.

Two of the consortium leads, Porter and Zingaro, have also written the classroom-ready version. Their book, Learn AI-Assisted Python Programming, reorders the introductory course around the skills the research elevates. They argue that the skills to write good software are evolving: problem decomposition, writing a specification, reading code, and testing matter more than they used to, while memorizing syntax and library details matters less. Two of the chapters teach students to read code, on the logic that if the assistant writes the code, the student’s job is to judge whether the assistant has done what the student intended. A full chapter covers testing and prompt engineering. Top-down design, the practice of breaking a large problem into smaller ones, runs across several chapters rather than appearing once. Debugging gets its own deep dive, a skill that think-aloud studies by Jacqueline Whalley and colleagues show novices tend to approach unsystematically. Porter and Zingaro grant that a future professional engineer should eventually learn to write code from scratch, but they argue it no longer makes sense as the starting place for most learners. That is a concrete answer to the camp that wants to ban the tools to protect the fundamentals: they are just sequenced differently and weighted toward the skills that matter most now, like reading code, testing, and breaking problems down, rather than writing syntax from memory.

What about large courses?

Class size does not change the goals, but it splits these design moves into two groups. Some work just as well with three hundred students as with thirty and cost the instructor no extra effort. Requiring an attempt before a student reaches for the tutor is a policy built into the assignment or the tool, not labor spent per student. Kumar’s guidance strategies were tested as exactly this kind of scalable intervention. The cognitive-engagement step is the same: Kazemitabaar’s Lead-and-Reveal shipped as a reusable task builder, so the prediction step runs for every student without the instructor touching it. The AI tutor itself has its strongest case here. Kumar frames LLM tutors as a response to growing class sizes where teacher presence is thin, CodeHelp was built for scalable support in large classes, and Stienstra’s students valued the tutor because it was always available, which a human teaching assistant in a 400-person course cannot be.

The moves that resist scale are both forms of assessment. Reading hundreds of process artifacts or grading prompts by hand does not work past a certain size. The realistic versions are to make the process artifact auto-gradeable or peer-reviewable, push it into teaching-assistant-led lab sections, assess prompts by outcome rather than by reading each one, and reserve the labor-heavy formats like oral checks for a rotating sample. The teaching of these skills scales fine, since a lecture is a lecture. The grading is what has to be redesigned.

Underneath this is the equity point. In a small class you can see who is drifting. In a large one, the student coasting on an illusion of competence is invisible, and that is the student Prather’s widening-gap study identifies as most at risk. The large-course adaptation is a triage logic: let the tools that scale support the median student, and aim scarce human attention, teaching-assistant hours, oral checks, early-alert outreach, at the students the tools are most likely to fail. Spreading human help evenly is the thing a big course cannot afford and the thing struggling students most need.

The opportunity

The strongest argument for bringing AI into programming education is not that the tools are impressive, though they are. It is that they force a reckoning with weaknesses the field has tolerated for decades: introductory courses where many students do not learn what we think they learn, and a persistent distance between what we teach and what graduates need. We can treat that reckoning as a threat to manage or as a chance to fix the course. The evidence says the fix is within reach, and that it depends less on the tools than on how we design around them.

A few things to consider trying next term:

• Replace one take-home assignment with an in-class task that assesses process, not just the finished program.

• Before students use an AI tutor, require them to attempt the problem and write down what they think the answer should be.

• Teach one short lesson on writing a specification for a model, including input and output types, and grade a prompt the way you would grade code.

• Pick a single cognitive-engagement move, such as having students predict what generated code will do before running it, and use it consistently for a unit.

References

• Denny, P., Prather, J., Becker, B. A., Finnie-Ansley, J., Hellas, A., Leinonen, J., Luxton-Reilly, A., Reeves, B. N., Santos, E. A., & Sarsa, S. (2024). Computing Education in the Era of Generative AI. Communications of the ACM. https://doi.org/10.1145/3624720

• Finnie-Ansley, J., Denny, P., Becker, B. A., Luxton-Reilly, A., & Prather, J. (2022). The Robots Are Coming: Exploring the Implications of OpenAI Codex on Introductory Programming. Australasian Computing Education Conference (ACE ’22). https://doi.org/10.1145/3511861.3511863

• Kazemitabaar, M., Chow, J., Ma, C. K. T., Ericson, B. J., Weintrop, D., & Grossman, T. (2023). Studying the Effect of AI Code Generators on Supporting Novice Learners in Introductory Programming. CHI ’23. https://doi.org/10.1145/3544548.3580919

• Prather, J., Reeves, B. N., Denny, P., Becker, B. A., Leinonen, J., Luxton-Reilly, A., Powell, G., Finnie-Ansley, J., & Santos, E. A. (2023). “It’s Weird That it Knows What I Want”: Usability and Interactions with Copilot for Novice Programmers. ACM Transactions on Computer-Human Interaction. arXiv:2304.02491

• Liffiton, M., Sheese, B., Savelka, J., & Denny, P. (2023). CodeHelp: Using Large Language Models with Guardrails for Scalable Support in Programming Classes. Koli Calling ’23. arXiv:2308.06921

• Kazemitabaar, M., Ye, R., Wang, X., Henley, A. Z., Denny, P., Craig, M., & Grossman, T. (2024). CodeAid: Evaluating a Classroom Deployment of an LLM-based Programming Assistant that Balances Student and Educator Needs. CHI ’24. arXiv:2401.11314

• Prather, J., Reeves, B. N., Leinonen, J., MacNeil, S., Randrianasolo, A. S., Becker, B. A., Kimmel, B., Wright, J., & Briggs, B. (2024). The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers. ICER. https://doi.org/10.1145/3632620.3671116

• Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A. E., Sarkar, A., Sellen, A., & Rintel, S. (2024). The Metacognitive Demands and Opportunities of Generative AI. CHI ’24. https://doi.org/10.1145/3613904.3642902

• Kumar, H., Musabirov, I., Reza, M., Shi, J., Wang, X., Williams, J. J., Kuzminykh, A., & Liut, M. (2024). Guiding Students in Using LLMs in Supported Learning Environments: Effects on Interaction Dynamics, Learner Performance, Confidence, and Trust. CSCW. https://doi.org/10.1145/3687038

• Kazemitabaar, M., Huang, O., Suh, S., Henley, A. Z., & Grossman, T. (2025). Exploring the Design Space of Cognitive Engagement Techniques with AI-Generated Code for Enhanced Learning. IUI ’25. https://doi.org/10.1145/3708359.3712104

• Stienstra, L., Mohamed, A., & Mohamed, M. (2025). Exploring GenAI as a Tutoring Tool: A Case Study in First-Year Computer Programming. ITiCSE ’25. https://doi.org/10.1145/3724363.3729060

• Shah, A., Chernova, A., Tomson, E., Porter, L., Griswold, W. G., & Soosai Raj, A. G. (2025). Students’ Use of GitHub Copilot for Working with Large Code Bases. SIGCSE ’25. https://doi.org/10.1145/3641554.3701800

• Garcia, R., & Craig, M. (2025). 20 Years Later: A Replication Study on Teaching CS1 Concepts. ACM Transactions on Computing Education. https://doi.org/10.1145/3730405

• Whalley, J., Settle, A., & Luxton-Reilly, A. (2023). A Think-Aloud Study of Novice Debugging. ACM Transactions on Computing Education. https://doi.org/10.1145/3589004

• Fisler, K. (2014). The Recurring Rainfall Problem. ICER ’14. https://doi.org/10.1145/2632320.2632346

• Watson, C., & Li, F. W. B. (2014). Failure Rates in Introductory Programming Revisited. ITiCSE ’14. https://doi.org/10.1145/2591708.2591749

• Bennedsen, J., & Caspersen, M. E. (2007). Failure Rates in Introductory Programming. ACM SIGCSE Bulletin. https://doi.org/10.1145/1272848.1272879

• Porter, L., & Zingaro, D. Learn AI-Assisted Python Programming, Second Edition: With GitHub Copilot and ChatGPT. Manning (Function). Kindle Edition. Available via IUCAT, Bloomington and Indianapolis campuses: ebook https://iucat.iu.edu/catalog/20590350, streaming audio https://iucat.iu.edu/catalog/21587005

• Novet, J., & Vanian, J. (2025, April 29). Satya Nadella says as much as 30% of Microsoft code is written by AI. CNBC.

• Consortium for Generative AI in CS Education, UC San Diego CREATE. Faculty leads: Leo Porter, Daniel Zingaro, Beth Simon. https://www.teachcswithai.org/

• Porter, L. Research on GenAI and Learning To Program [Video]. YouTube.

Picture the chair. Cape on, and your stylist swivels a tablet toward you like it’s a tarot reading. “Let’s see what the AI says about the shape of your face.” You upload a photo. A little wheel spins. Then the verdict: heart-shaped, and here are the three cuts that will change your life. It hasn’t happened to you yet, but it’s coming, because the app that does this already exists. The app is mine. And the AI reading your face is not what I told you it was.

I built StyleMatch the way I now build most things, by describing what I wanted in plain English and letting an AI scaffold it. It’s a consultation tool for stylists. Upload a client photo. Get a face-shape read. Browse curated styles. Drop your favorites onto a “Look Board.” Generate a try-on preview that pastes a new hairstyle onto the client’s actual picture so they can see it before anyone picks up scissors. It came together fast. It looked great. I almost shipped it on vibes alone.

Then I did the thing I taught myself to do back when I was dragging Real Housewives around a screen. I asked an AI to open the hood and explain it to me, line by line.

Two stories live in there. One I’m proud of. One I’m confessing.

The part that’s real, and kind of great

The try-on feature works. You upload a photo, pick a style, and the image model genuinely renders it onto the person. The thing I’m proud of isn’t that it works. It’s the instructions I gave it. They say, in plain terms: change only the hair. Keep the person’s face. Keep their skin tone. Keep their ethnicity, their pose, their background. Touch nothing but the hairstyle.

If you don’t work in this space, that might read as obvious. It is not. These tools have a documented habit of “beautifying” people by lightening skin and softening or erasing features that don’t match the training data’s idea of a default face. I knew that going in, so I wrote the guardrails: a Black woman trying on a bob comes back as herself with a bob, not as a stranger. That’s not a flashy feature. It’s the part nobody notices if I did it right, which is exactly why I wanted it right.

Was it perfect? No. I ran my own photo through, and it kept everything I told it to keep, then aged me anyway. A few fine lines I don’t remember authorizing, enough that I bought a bottle of collagen within the hour. I would show you, but vanity prevails. The guardrails held where it mattered, on skin tone and features and the things these models get wrong in ways that actually harm people. They just didn’t stop the model from having opinions about my face.

But the hair looked great. I came away with a real sense of what a new cut and color would do, and a decent image, hairwise, to hand my next stylist. The feature did the exact job I built it for.

The part that’s fake

Now the face-shape analysis. The headline feature. The thing the stylist points at when she says the AI read your face.

It does not read your face. I know, because I wrote it. Or rather, I described it, let the AI fill it in, and didn’t check.

What it actually does is take the photo’s web address, the link, and run a little math on the characters in that link. It adds them up, more or less, and maps the result to one of seven face shapes. Same link in, same shape out, every single time. The actual face, the pixels, the bone structure, the thing I told the user it analyzed, never enters the calculation. Upload a photo of a lamp under the same link and you’d get the same confident “heart-shaped” verdict.

It wasn’t malice. It was a placeholder, the kind a coding tool drops in to make the screen work while the real thing gets sorted out later. My own code comments say so, in writing, to me. Later just hasn’t come. And here’s the part I have to own: StyleMatch is live right now. Testing mode, nobody’s being charged. But the landing page says “7 face shapes, analyzed by AI,” and today that sentence is misleading. I have to take ownership of that.

Why this is the whole story of AI right now

It usually comes down to the same gap. The distance between what an AI product claims and what it does is almost always invisible, because the marketing sits on top of the code, and the code is the part nobody reads. Including, it turns out, the person who built it. I got the spinning wheel and the confident verdict and almost believed my own app, because I never saw the comment that said “placeholder.” It took another AI, Claude Code, reading the thing back to me in a format I could understand before I understood what I’d shipped. An AI caught the AI. The irony is that I am still relying on AI to explain to me whether something works as I intended.

What makes StyleMatch a clean teaching case is that you don’t have to pick a side. Honesty and hype are sharing one app, one screen, two features. The try-on is real and built with care. The face read is a magic trick. I shipped both. Both say “AI.”

And here’s where it stops being about hairstyles. Think about how we adopt tools in education. A vendor runs a demo. The demo is gorgeous. There’s a landing page with a number on it: “trusted by,” “analyzed by,” “personalized for.” A committee meets, and real vetting happens. In a 2025 EDUCAUSE poll of higher-ed institutions, data security and legal compliance were the top concerns, each named by 86 percent. That’s the right instinct. But look at what that review is built to catch. It asks whether a tool is safe and legal, not whether the AI inside it does what the label says. Algorithmic bias came in last, at 30 percent. And the people most likely to notice when a tool doesn’t actually work, the teaching and learning staff, were in the room only 58 percent of the time, behind IT and security. So a tool can clear procurement, sit inside every privacy rule, and still be a hash function in a lab coat. Nobody lied. The review just wasn’t pointed at that question. I would know. I built one that would pass, and I only caught it because I went looking.

So

Next time something tells you it analyzed you, ask it to show its work. Ask what it looked at. Ask what happens if you feed it the same thing twice, or feed it nonsense and see if it blinks. I’m asking it of my own app now. The honest version of StyleMatch is completely possible to build. A real model can read a real face. I just haven’t built it yet, and the landing page is already making the promise. That one’s on me.

Me:

Now, my sister loves me. “Nice” was the kind thing to say. But I’ve spent fifteen-plus years studying how people learn, and I just got a one-line teaching evaluation from someone who wants me to win. Not at all.

I was sitting with that when I logged into a class on faceless content. The instructor is a creator who built his audience on reality TV recaps, the housewives, the messy reunions, the text-message screenshots. Richie Skye. The reasons I chose to take his course are:

1. He’s very successful

2. He’s thoughtful in his commentary. As you may have guessed by now, I also watch some of the housewives recaps. He is one of the few content creators who seems to remember that, no matter how wild the women’s antics are on the show, they are still humans, with feelings, and families that care for them. He often reminds his viewers of this point when they discuss the show in the comments.

3. He encourages his viewers to be more than passive watchers, but get out and live life and embrace an entrepreneurial spirit. Specifically, I remember him saying his goal for participants to surpass his own success, viewing their achievements as the ultimate compliment. He offers free, low cost, and moderately priced instructional content in addition to the housewives content. You can take a group course or work with him directly: https://stan.store/djrichieskye/p/book-a-11-breakthrough-strategy-call-with-me

On paper we have nothing in common. He talks about super chats. I talk about situational interest and active learning. But he had a lot to teach me that nobody in my professional world had managed to teach me. About twenty minutes in, he said the thing that has been rattling around my head ever since, and it had nothing to do with money.

He called it the black licorice test.

Here’s the idea. Most people don’t like black licorice. If he stood in front of a crowd and asked who loves it, maybe one hand goes up. And his point was: that one person is the whole game. You don’t make content for the room. You make it for the person who came for exactly the flavor you make. Everyone else can wander off. He said it plainly: “I don’t care about those people. I care about the people I need to serve with my content. Who is it for?”

I laughed, because I teach this. I just teach it in a costume.

Good teachers do this without thinking. Before they explain anything, they ask who is actually listening: what those people already know, what they’re confused about, what would bore them, what would move them. My field has a duller name for it, audience analysis, and it’s the first move in instructional design. It’s also the step everybody nods along to in the workshop and quietly skips the moment they sit down to build the real thing.

And I had skipped it, despite how many times I warned against it! Of course I had. I made my videos for a person who already understands AI in education, because that person lives in my head and never asks me to slow down. I’d say prompt engineering, or warn about hallucinations, and never stop to notice that some viewers had no idea what either word meant. That person is me. My sister is not that person, and my sister is far closer to my actual audience than I am.

This has a name: the curse of knowledge. Once you know a thing deeply, you genuinely cannot remember what it felt like to not know it. You say “it’s just a model” or “you can prompt it to” and you don’t notice the four assumptions stacked under that sentence. Expertise makes you a worse explainer by default. You have to fight your way back to clarity on purpose.

The content creators don’t have that problem, and here’s the uncomfortable reason: they get paid only when they’re understood. If the recap doesn’t land, nobody watches, nobody chats, nobody comes back. The feedback is brutal and immediate. So they’ve built an entire craft around one specific viewer and one clear job per video. They title things so plainly it’s almost rude. They put the subject right there on the screen. No throat-clearing, no preamble, no “before we get into it.” Just the thing, made for the person who wanted the thing.

I want to be careful here, because this is where it usually goes wrong. Clarity is not dumbing down. Those are different moves, and people collapse them together so they have an excuse to stay obscure. The test I hold myself to is this: a person with no background should walk away understanding the core idea, and an expert should find it honest rather than cheapened. Both at once. That’s harder than writing for experts, not easier. Plain language is a discipline, not a shortcut. The clear version of an idea is almost always the hardest one to write, not the easiest.

But for me it reads as something closer to a teaching principle. The tools can draft, summarize, restructure, and they’re genuinely good at it. What they can’t do is be the person the learner trusts. And that loops straight back to the audience. People aren’t only here for information, they can get information anywhere. They’re here for someone who remembers what it was like not to know, and explains it anyway. Accessibility was never only about the words being simple. It’s about a real someone deciding you’re worth understanding, and building the on-ramp so you can.

So here’s what I’m changing. I’m not making videos for the version of me that already gets it. I’m making them for my sister, who is smart and curious and busy and owes me no patience. One idea per video. The point in the first ten seconds, not the third minute. The title says what it is. And before I publish, I’m sending it to her, and the only question that counts is the one I should have asked first.

Did you understand it?

Here is the part that got me: it is mostly right. It knows I will stop for a raw or fried oyster. It knows there is a Prince-shaped hole in my listening history. It knows I spend my days thinking about how people learn. It also believes, with total calm, that I live in Ashburn, Virginia.

I do not live in Ashburn, Virginia. My internet just visits there sometimes.

Let me give the machine its due, because the hits are good. One card told me to draft my Phase 1 abstract for the AECT Design Showcase, with a picture of me sketching learning concepts at a sunny desk. AECT. The Association for Educational Communications and Technology. That is not a lucky guess. That is my actual field. I have held offices in and presented at AECT.

Another card pitched me the new Siri AI beta for my “automation workflow.” A third nudged me toward gathering STEM layout ideas, leaning on what it called my “background in building learning kits.” Not wrong about that either. The illustrated me in these cards is always mid-project, always a little busy, always making something. If you wanted to flatter an instructional designer, this is how you would do it.

Then there is the Prince of it all. One card invited me to stream “Stone,” (you’re welcome), a newly surfaced soul track from the Paisley Park vault. I tapped the little button that allows for questions. Mine was, Why this? (But I really wanted to know what information it used from my use of Google tools to suggest this).

Here is what it told me, verbatim:

It looks like this story was suggested because of your interest in classic artists like Prince. It appears your search history shows a pattern of exploring music discographies, and this story about the newly released track “Stone” seemed like a great fit to share with you, especially following the recent Prince Celebration in Minneapolis.

Reader, it had my number and I felt seen. The affinity is real. The discography rabbit holes are real. I sat for a second, a little exposed, a little charmed.

And here is the honest part. None of this surprises me. I know how these systems work, and I know what I have typed into a search bar at one in the morning. The surprise was never going to be the point. What I cannot quite settle is how I feel about seeing myself this specific, my field, my conference, my late-night Prince spiral, all of it rendered back to me as a tidy watercolor profile. Knowing the trick does not stop it from being shopping.

So I pushed the same button on a different card: fresh seafood platters at Fords Fish Shack. Why this one?

It looks like this story was featured because it aligns with your known interests in fried fish and fresh oysters. It seems your profile also highlights that you are currently based in Ashburn, Virginia, and we wanted to share a local dining option that would be perfect for the warm June weather.

Fried fish and fresh oysters: yes, guilty, keep them coming. (I would never where that dress though). Currently based in Ashburn, Virginia: absolutely not. And there it was, loud and wrong, sitting right next to the thing it had nailed.

Once I saw Virginia, I saw it everywhere. The aircraft card sent me to an Innovations in Flight showcase a "quick drive from Ashburn." The pollinator festival was in Leesburg. Fords Fish Shack is an Ashburn spot. Every outing the app planned for my Saturday was forty minutes from a house I have never lived in.

Here is what I think happened. My internet, for reasons involving a VPN and the routing decisions of my service provider, sometimes shows up in Northern Virginia. Dreambeans took that one signal, an IP address, and quietly promoted it to a fact about my life. Then it fused that fact with the real me, the one who loves Prince and oysters and good pedagogy, and served the blend back as a single confident person. Half of that person is me. The other half is a server rack outside Dulles.

I study how people learn, and a chunk of that work now means watching how machines decide what to put in front of us. So I cannot file this under cute mixup. The interest model is sharp. That is exactly why the Virginia error is worth sitting with, not in spite of it.

Notice the tone. When the app explained Prince, it was right, and it sounded certain. When it explained the fish shack, it was wrong about where I live, and it sounded exactly as certain. Same even voice, same tidy sentence, no flicker of doubt between a strong inference and a bad fact. The “why” button is a real improvement, and I want more of it, not less. But an explanation that cannot tell me how sure it is on each piece is only doing half the job.

That gap is the thing I would flag for anyone building or buying these systems for classrooms, which is the room I actually work in. A recommendation engine that knows a student’s interests is useful. A recommendation engine that confuses a signal with a self, and reports both in the same confident register, is the one that quietly sorts people into the wrong box and never says it happened.

I am keeping the Prince track. It is good, and the app was right to send it.

I am still not in Virginia. Not sure I’m going to correct it, or see if it will eventually self correct.

That, more or less, is the global conversation about AI safety right now.

Everybody is saying “AI risk.” Brussels says it. Silicon Valley says it. Your dean said it in a meeting last week and you nodded. And almost none of us are checking whether we mean the same thing. One framework’s “safety concern” is another’s “technical limitation.” The words sound shared. The meanings are not. And what should bother you more than it does: you cannot coordinate a response to a borderless technology when the people at the table are using the same vocabulary to describe different fears. The fight isn’t even productive. It’s just loud.

So a team at MIT FutureTech sat down and did the thing nobody wanted to do, which is read everything. They built the AI Risk Repository, and to get there they went through 17,288 articles, pulled out 74 separate risk frameworks, and sorted the whole mess into a database of 1,725 individual risks. Then they organized it two ways. One taxonomy maps where a risk comes from (the who, the when, the why).

The other maps what the risk actually does to people, sorting every harm into seven domains of consequence: discrimination and toxicity, privacy and security, misinformation, malicious actors and misuse, human-computer interaction, socioeconomic and environmental harm, and AI system safety. Twenty-four subdomains sit under those seven. Think of the whole thing as the glossary the reunion never had. Here are the five key points.

The “jingle-jangle” problem, or how one word starts a war

The researchers borrow a pair of terms from psychology that I now want tattooed somewhere visible: the jingle fallacy and the jangle fallacy. A jingle fallacy is when one word gets stretched across totally different things. A jangle fallacy is the reverse, when a bunch of different words all turn out to mean the same underlying thing. Both of them are quietly wrecking our ability to govern this technology.

Take “privacy.” In one framework it means a model coughing up sensitive data it memorized during training. In another it means a government watching everyone all the time. Those are not the same problem. One is a bug. One is a human rights crisis. When the same word covers a coding error and a surveillance state, writing rules that travel across borders becomes close to impossible. The U.S.-EU Trade and Technology Council put it plainly in its Trustworthy AI roadmap: “Shared terminologies and taxonomies are essential for operationalizing trustworthy AI and risk management in an interoperable fashion.” Translation: pick a glossary, or keep fighting about nothing.

We’re judging the plate and ignoring the prep

Here’s where the timing data gets uncomfortable. When you sort the frameworks by when a risk shows up, 62% of them focus on what happens after a model is already out in the world. Only 13% deal with the risks baked in during development and training.

Think about Top Chef. The judges taste the plate at the end, and that’s the moment everyone remembers: the verdict, the “please pack your knives and go.” But ask anyone who’s survived a season and they’ll tell you the dish was won or lost hours earlier, back in the Whole Foods aisles and the mise en place. By the time it hits the pass, the outcome is mostly cooked, literally. That’s the development phase. It’s where the choices get made that you can’t un-make later, where risk gets welded into the architecture before the public ever takes a bite. The researchers say it cleaner than I can: “excessive focus on post-deployment risks may blind us to threats from development itself.” By the time a system is live and you’re reacting to harm, the judges are already tasting. As an instructional designer I feel this in my bones, because the same thing is true of a course. The damage usually isn’t in the delivery. It’s in the design.

It’s not just the bot. It’s the boss.

There’s a story we like to tell about AI risk, and it stars a rogue machine. Very cinematic. Mostly wrong. When the Repository looks at who is responsible for a given risk, the blame splits almost evenly: 42% of risks trace to the AI systems themselves, and 38% trace to human decisions and actions. The machine and the management, roughly neck and neck. The Repository asks a companion question too, which is whether anyone meant it, and the answer is just as split: 35% of mapped risks are intentional, 37% are not, and the rest are unclear. So the deliberate villain we love to picture, the hacker, the authoritarian, the scammer, is barely a third of the story. Most of the danger is harm nobody set out to cause.

There is another number that should be noted as well. The remaining 21% lands in an “Other” bucket, and what’s in that bucket is risk that nobody caused alone. It’s emergent. It shows up specifically in the interaction between the human and the system, the way some arguments only exist because two particular people are in the same room. You can’t pin it on the bot. You can’t pin it on the boss. It only happens when they collide. Which means governing AI isn’t just about better code or better oversight. It’s about governing a relationship, and relationships are exactly the thing institutions are worst at.

The frontiers nobody’s mapping yet

When you sort risks by what they actually affect, two categories are almost missing entirely. AI welfare and rights shows up in 3% of frameworks. Multi-agent risks show up in 7%. Two of the most consequential questions about where this is headed, and the field has barely written them down.

I get why. Today they read as philosophy-seminar material, the stuff you debate at 11pm and walk away from by morning. Not forget, exactly. More like what learning scientists call incubation: a problem too big to crack on the spot, so you set it down and let it simmer in the back of your mind while you gather more information. That can be the smart move, as long as you actually come back to it. Multi-agent risk in particular stops being abstract the second AI systems start talking to each other instead of just to us. The researchers flag things like “selection pressures” and “collusion,” patterns that emerge when independent agents interact with no human in the loop, and that can cascade into failures nobody designed and nobody saw coming. We are building toward a world of autonomous systems negotiating with other autonomous systems, and our risk maps mostly stop at the edge of that ocean and write “here be dragons.” Except we’re already sailing.

Naming the danger is how you clear room to build

Here’s the turn, and it’s where I’d push back on anyone who calls this kind of work a buzzkill. The MIT team argues that mapping risk carefully is what frees people to innovate. When you make risks explicit and categorical, you also define the safe harbors, the parts of the design space where developers can move fast precisely because everyone understands where the edges are.

History backs this up. Naming the risk of discrimination is what produced fairness-aware machine learning. Worrying out loud about alignment is what produced frontier model evaluations. The risk didn’t kill the progress. The risk was the assignment. A clear map lets investors and builders stop guessing about uncertainty, which, as the researchers put it, “can accelerate beneficial AI deployment... protecting society from AI harms while enabling the realization of AI’s transformative potential.” You don’t drive faster by ripping out the guardrails. You drive faster because they’re there.

We’re building the map mid-flight

The Repository isn’t a finished document. It’s a living database, updated as the technology moves, which is the only honest way to do it when 93% of the frameworks it covers were published after 2020 and a lot of them are still preprints. We are drawing the map while the plane is already in the air, and pretending otherwise helps no one.

If we’re going to govern systems that are getting more multimodal and more autonomous by the quarter, the first job isn’t technical. It’s linguistic. We have to agree on what we’re naming before we can agree on what to do about it. A shared vocabulary is the bridge between the lab and the legislature, and right now it’s half-built.

And since most of you reading this teach, or design what other people learn from, this lands closer to home than it looks. The vocabulary problem is already in your syllabus. “AI use” means a banned shortcut on one professor’s assignment, a required skill on the one next door, and whatever got them through last semester to the student. That’s a jingle fallacy with a grade attached. The pre-deployment lesson holds too: the time to decide how AI belongs in a course is while you’re designing it, not in week ten when something lands in your inbox you can’t quite source. And that 21% nobody causes alone is every student-and-chatbot pair in the room, turning in work neither the kid nor the tool would have made on its own. You don’t govern that with a ban. You design for it, or you get surprised by it.

So I’ll leave you where the reunion left those two women, except I want us to do better than they did. Can we actually “align” AI if we’ve never once aligned our own understanding of what we’re afraid of?

There’s a particular kind of misery in being the new person. You’ve got the job. You’re qualified. Somebody decided you were worth hiring. And yet here you are in week one, staring at a system you don’t understand, trying to figure out why this one function takes fifteen arguments, none of which are named anything a human would name them. You can’t find the light switch. You know there’s a light switch. Everyone else walks into this room and the lights just come on. So you do the thing you’ve been doing all week: you go ask the one person who knows, again, and you watch a flicker of “didn’t I just explain this” cross their face before they’re kind about it.

I think about this scene a lot, because it’s not really an engineering problem. It’s a learning problem wearing an engineering costume. And it happens everywhere, not just in code. New teachers, new nurses, new line cooks. The expertise they need is sitting in someone else’s head, and most teams have no real system for getting it out.

The thing experts can’t quite say

Here’s what learning research figured out a while ago. The most important knowledge an expert has is the knowledge they can’t see anymore.

Think about how a surgeon actually learns, or a chef, or a teacher in their first semester. Nobody becomes competent by reading the handbook. They become competent by standing next to someone experienced while that person makes real decisions in real situations. The judgment that matters is mostly invisible. The senior architect knows why the legacy pattern exists, but they stopped consciously knowing that they know it years ago. Ask them and you’ll get “oh, that’s just how we do it here,” which is the least helpful sentence in the English language.

Learning researchers call the fix cognitive apprenticeship. The idea is simple to say and hard to do: because thinking is invisible, good teaching has to drag it into the open on purpose. You reason out loud. You coach while the novice tries. You hold up the hard parts, then take the supports away as they get stronger. A line cook gets some of this for free, because the work happens out in the open where you can watch the hands move. The architect, the professor, the senior anybody, has to do it on purpose. It’s slow. Outside of places built for it, like residencies and student teaching, almost nobody schedules time for it.

So mostly we don’t. We pay for it the expensive way instead, one interruption at a time. Every time the new hire pulls a senior out of deep work to ask why, that interruption is the apprenticeship happening. Unscheduled, unrepeatable, gone the second the conversation ends.

Two curricula

It helps to notice that every workplace is secretly teaching two different things at once.

The first curriculum is the system: the codebase, the procedures, the rules, and the reasons they all quietly became what they are. The second is the culture: what good judgment looks like here, what counts as a strong argument, which corners this team cuts on purpose and which ones it would never. The first you can write down. The second you mostly can’t, which is why it gets transmitted by being in the room.

Hold onto that split, because it’s the whole story. AI is getting remarkably good at teaching the first curriculum. The question this essay is actually about is what happens to the second.

What the machine changes

Let me be concrete, because the news here is genuinely good before it gets complicated.

Picture a new engineer who wants to know why the payment system retries a failed charge exactly three times. Not two, not five. Three. Twenty years ago, finding out meant locating the person who wrote it, assuming they still worked there, assuming they remembered. Today an agent can trace the commit history, surface the bug report from a few years back, and explain that a banking partner started rate-limiting after the third attempt, so somebody capped it and moved on. The new engineer gets the answer in a minute, and nobody got pulled out of deep work to give it.

All of this depends on the trail being there. When the commits are clean and the bug report got filed, an agent can reconstruct a genuinely good explanation. When the trail has gone cold, it will still hand you one, just confident and possibly fiction. But where the record exists, these tools are surprisingly good at rebuilding the kind of explanation that used to require interrupting a senior engineer, and they do it without sighing.

That’s the system curriculum, taught on demand. The new hire can pull on it themselves, at their own pace, at eleven at night, without the social tax of looking lost in front of the people who just hired them.

Who used to get the apprenticeship

And here is the part that is interesting to me, because it’s an equity story hiding inside a productivity one.

Informal coaching has always been handed out unevenly. The junior who happens to sit near the senior gets the apprenticeship. The one who feels safe enough to keep asking gets it. The confident one, the one who reminds the senior of themselves, the one who got invited for coffee. Everyone else learns to fake it and prays nobody notices. An agent that answers the same question for the tenth time without a flicker of impatience doesn’t fix that on its own. But it does something real. It hands a door to the people who never got pulled aside.

That’s not nothing. For a lot of newcomers, that’s the difference between belonging and bluffing.

What the machine can’t reach

Here’s where the two curricula come apart.

If teaching were only information transfer, we’d have solved onboarding with good documentation decades ago. We didn’t, and not because the docs were bad. Even teams with excellent documentation still can’t seem to transmit judgment. That’s the tell. The hard part was never the facts. It was everything around the facts.

The culture curriculum doesn’t move through answers. It moves through proximity. You learn what contextual good judgment looks like by watching people argue about a tradeoff, change their minds, defer to the person who’s been burned by this exact decision before. You absorb the team’s taste by being near it while it operates. An agent can tell you what the team decided. Record every meeting and it could even replay the argument they had getting there. What it can’t do is make you a participant in it: teach you the tone, the timing, when to push and when to let a thing go. It can show you the room. It can’t put you in it.

So here’s my worry, and I’m holding it as a real question, not a verdict. A new hire who learns from an agent can get fast. Genuinely, impressively fast. They can find every light switch in the house. But they may never learn to live in it. They get the answers without the arguments behind them, the system without the culture. Fluent in the codebase, and still a stranger to the people who wrote it.

That’s not a reason to keep the tools away from anyone. I’d hand one to every new hire tomorrow. It’s a reason to be honest about which parts of apprenticeship were always meant to be slow. Asking the same question twice? Automate that, please, free everyone. But sitting in a messy decision with people who’ve made it a hundred times was never inefficiency. That was the actual teaching, disguised as a meeting.

And this isn’t only about new engineers. It’s the graduate student, the medical resident, the first-year teacher, the freshly promoted manager: anyone learning a craft inside a community whose real lessons were never written down. The system curriculum is the part we always knew how to hand off. The culture curriculum only ever transferred by being there.

Which points at one thing organizations might actually do. If an agent becomes the first mentor every newcomer meets, then the human mentoring has to get more deliberate, not less. Someone has to protect the rooms where people still reason out loud together, and make sure the new person is in them, because that’s the curriculum the machine can’t teach and nobody will think to schedule.

The question isn’t whether the machine can teach the new person. It clearly can teach them plenty. It’s whether a workplace will notice when the system curriculum quietly goes automatic and the culture curriculum, the one nobody scheduled in the first place, simply stops being taught, or not taught well enough to help a new person adapt to the environment. We’ll know the lights came on. What’s harder to notice is whether anyone ever learned to live in the house.

It’s a Sunday and I am, regrettably, watching The Biggest Loser. Not the new one. The original. The one where contestants disappear to a ranch in California, cry through a montage, and emerge a hundred pounds lighter in eighteen weeks. The reveal is theatrical. The music swells. Someone’s mother gasps.

Then they go home. And follow-up research years later found their metabolisms had collapsed and almost all of them had regained the weight. The transformation, it turned out, was a performance.

Now flip the channel. Top Chef. A contestant has ninety seconds to explain her plate to Tom Colicchio and you can see her doing the math: do I tell him I meant this, or do I tell him the truth? She tells him the truth. He nods. Next week she does it again, only this time the dish lands.

These two shows are the cleanest illustration I know of the difference between knowledge and transformation. And once you see it, you cannot un-see it in your own teaching.

Knowledge is cheap now. Transformation is the whole course.

I’ve been thinking about this because of Vishen Lakhiani’s five-praxis model. I’ve been on a steady diet of growth-mindset books and podcasts lately, and his framing is the one that keeps doing work on how I think about course design. Transformation, the model argues, isn’t something that happens to you. It’s something you practice. The five praxes:

1. Critical Reflection

2. Critical Study and Writing

3. Social Discourse

4. Rate of Application

5. Altered States

Read those out loud. Now picture Top Chef. Every single one is on screen, every single episode.

The Judges’ Table is critical reflection at industrial scale. The chefs read each other’s plates in the stew room (critical study). They argue, defend, concede (social discourse). The Quickfire forces the technique to appear in your hands within an hour (rate of application). And the long Restaurant Wars night, when everyone is exhausted and half-delirious and the menu finally clicks at 2am, is the altered state nobody planned for but everyone needs.

Now picture The Biggest Loser. A trainer screams. A scale appears. A montage runs. The “reflection” is a confessional booth where the contestant cries about their dead father. The “study” is a nutritionist holding up two plates. The “discourse” is a fight on the yellow team. The “application” only exists inside the ranch. And the “altered state” is the reveal, where the contestant walks through a paper screen in front of their family and the camera lingers on the mother gasping.

One of these is praxis. The other is content.

The CS classroom is where this argument gets stress-tested

I work in a computing school, so I’ll come back to my home discipline, but the argument generalizes to any course that has a craft inside it. Computer science makes this distinction impossible to ignore, because the knowledge work in CS has been substantially automated, and the artifacts students produce (code that compiles, tests that pass, an assignment marked correct) can look identical whether they understood the work or not. Some of us have started asking students to submit pseudocode alongside their code, forcing them to articulate the rationale and document the knowledge they’re claiming. It helps. It doesn’t fully close the gap, because most rubrics were calibrated to artifacts, not to the process that should have produced them. A first-year student can still prompt a chatbot for a working binary search tree at 2am and submit it before sunrise. The tree compiles. The tests pass. The pseudocode reads well. And the student is exactly who they were the day before.

This is The Biggest Loser model of education. We’ve been running it for years and only the receipts have changed.

What Top Chef teaches us is that transformation requires the praxes to be wired into the structure of the work itself, not bolted on as enrichment. Here’s what that looks like in a course. The examples are computing-shaped because that’s where I teach, but the artifact is the variable, not the praxis.

Critical reflection stops being “submit a one-page reflection” (the educational equivalent of “how do you FEEL?”) and becomes a debugging journal with a required format: hypothesis, test, result, revised mental model. By week eight, students start citing their own past entries. They start arguing with themselves. That’s the move.

Critical study and writing stops being assigned readings nobody finishes and becomes codebase annotation. Pick a real open source repo. Pick one module under 500 lines. Annotate every design decision, every assumption, every “why this and not the obvious alternative.” Submit the file. The reading is now visible because the writing is now visible. Substitute “annotate this primary source” or “mark up this contract” or “annotate this lesson plan” for your own discipline. Same move.

Social discourse stops being discussion-board theater (you know the one: “Great point, Kayla, I also believe...”) and becomes adversarial peer review. Every submission requires at least one substantive critique from a peer. No rubber stamps. Students grade each other on review quality, not work quality. The stew room comes to the classroom.

Rate of application is where most courses quietly die. We introduce a concept in week three and assess it on a midterm in week seven. The Quickfire approach is better: every new concept appears in the student’s own work within 24 hours, in a tiny artifact they’d actually use. A log parser, a file renamer, a one-page argument, a single annotated slide. The skill must show up in your hands while the lecture is still warm.

Altered states are the praxis we’re worst at honoring, partly because the assignment patterns we use make them impossible. Students start the night before, submit at the deadline, and there is no morning-after. If we want the subconscious to get a turn, we have to design the time into the assignment, not hope students will find it on their own. That means staged deadlines (draft due Tuesday, final due Friday, with the gap enforced). It means final submissions that have to be timestamped at least 24 hours after the last commit. It means in-class sessions where the assignment is to step away from the screen and journal about a stuck problem. The chefs on Top Chef don’t have late-night menu breakthroughs because they decided to be reflective. They have them because the show structured the time. The classroom equivalent has to do the same work.

Where the AI shows up

I’m not going to pretend AI doesn’t change the calculus here, because it does. But it changes it in one specific way: it accelerates knowledge work and threatens transformation work.

The praxis model is actually a great filter for deciding where AI belongs in your course. Anywhere the work is about acquiring or verifying information (syntax, definitions, “what does this error mean,” “what’s the standard form of this citation”) AI is fine. Anywhere the work IS the transformation (the first implementation of a new concept, the reflection writing, the peer review, the long sit with a hard problem), AI displaces the very thing you were trying to teach.

This is the Biggest Loser logic again. AI-generated work is the paper-screen reveal. The student walks through it and the grading software gasps. Their classmate’s mother is moved. The student goes home and six weeks later cannot debug a basic loop, or argue a basic claim, or summarize a basic source.

A syllabus written through the praxis lens names this directly, in writing, in the assignment specs. AI-permitted: looking things up, generating test cases after you’ve written the function, debugging support after a documented attempt, explaining unfamiliar material. AI-free: first implementation of any new concept, peer review (you are the reviewer), reflection writing, the final project’s core argument. You can argue with my specific list. The principle holds.

On the labor question

A fair pushback: this is more work for faculty. It is, up front. The praxis-shaped rubric, the assignment redesign, the syllabus rewrite, all of it costs hours the first time. But it isn’t more grading work, because the assessment shifts from evaluating finished artifacts (which now requires forensic detective work) to evaluating process artifacts (the journal, the pseudocode, the peer review, the timestamp gap). The first kind of grading was already breaking. The second kind is the work we should have been doing all along. Pick one praxis. Redesign one assignment. See what changes. The rest can wait.

The closing turn

Reality TV figured something out that higher ed mostly hasn’t. The shows that produce real transformation are the ones that subject their participants to the praxes over time, with stakes, with peers, with reflection built into the format. The shows that produce the appearance of transformation are the ones that skip to the reveal.

Knowledge acquisition was always a means. Transformation was always the point. We were able to confuse the two for a long time because the knowledge work was effortful enough to feel like learning. It isn’t anymore. Which means the praxes are no longer optional. They’re the syllabus.

Three things to try this term.

1. Pick one assignment and rewrite it so that one of the five praxes is the only sensible way to complete it. Just one. See what changes.

2. Name your AI-permitted and AI-free zones in writing, in the syllabus, by praxis. Don’t leave it to vibes.

3. Build in one walk-away. Sleep-on-it submission rule, mandatory break protocol, something. The subconscious wants in.

Tom would nod. The Biggest Loser would cut to commercial.

It’s Sunday night and your boss wants a self-review by Friday. Or it’s December and you’re updating your résumé, staring at the year and trying to remember what you actually did. Or your therapist asks how the last six weeks have been and you blank, because they live in a fog somewhere between school pickup and a work Slack you’d rather forget.

If you’re like most people I know, the honest answer to “what have you been up to” is everywhere and nowhere. Some of it’s in Slack. Some of it’s in your head, vaguely. A lot of it never got written down at all, because helping your sister move or sitting up with a sick kid doesn’t generate a paper trail. It generates a knot in your shoulder.

Here’s the thing: your calendar already holds most of what you need. You just have to set it up to give it back to you. Ten minutes now. A few seconds per entry going forward. The payoff hits at the worst possible time of year, which is exactly when you need it.

Why your calendar and not your brain

Email is a swamp. Try to reconstruct what you did last quarter from your sent folder and you’ll lose a Saturday. Your résumé only captures the things that already have a name: promotions, certifications, the trip you took. It doesn’t capture the work between them.

Your calendar is different. Every entry has a date, a length, the people involved, and a title you control. It’s already structured. It’s already searchable. It’s already syncing to three of your devices.

The catch is that most of us name calendar entries for our right-now selves (”Dr appt 2pm”) instead of our future selves, who six months from now will be trying to figure out whether that appointment was the dentist, the dermatologist, or the one with the orthopedist about the knee.

Name the entry for your future self. That’s the whole post.

Five habits that pay off later

1. Use a naming convention. Pick short prefixes that tell you the category at a glance. Mine are Work, Health, Money, Family, Home. So an entry reads “Health: PT for shoulder” instead of “appt.” Pick prefixes that match the questions you’ll actually need to answer later: tax time, performance review, custody log, whatever your life looks like.

2. Use color categories. Most calendar apps let you assign colors. Right-click, pick one, move on. Six is plenty. When you want to see everywhere I spent money on the house this year or every time I drove the kids somewhere, the colors do the work for you.

3. Log the invisible stuff as ghost entries. This is the move that changes everything. After the hour you spent on the phone fighting your insurance company, add a 15-minute calendar entry, in the past, titled “Money: Aetna appeal call, ~1 hr.” You’re not lying to your calendar. You’re telling the truth to your future self. The mental tax of that call was real. The receipt should be too.

4. Put the outcome in the title. When something produces a result, drop it in. “Home: HVAC tech replaced capacitor, $340” tells you more in five years than “HVAC guy” ever will. Future you is going to thank present you when the warranty conversation comes up.

5. Block your own time and label it. A two-hour block called “Work: draft Q3 deck” is documentation. A two-hour block called “focus” is not. Same time on your calendar, completely different evidence trail.

You won’t do all five perfectly. I don’t. The goal is to make the calendar slightly more truthful than it was last year, not turn yourself into your own surveillance state.

The privacy part, because (I hope) Severance is fiction

Your calendar is not a diary. It’s a record on a server you don’t own. Your employer can see the work one. Your phone backup syncs the personal one. Anything you put in plain text could, in theory, be subpoenaed, hacked, shoulder-surfed during a screen share, or read aloud by Siri when your kid grabs your phone at the coffee shop.

So a few rules I follow, and you might want to:

Medical stuff stays vague. “Health: appt 2pm” is enough. You don’t need “therapy session, discussing the anxiety meds” sitting in cleartext. You don’t need “Dr. Patel, mammogram follow-up” as a calendar title that previews on your lock screen at brunch. Future you can remember what the appointment was for without the clinical breadcrumb. “Health: specialist visit” carries the same weight on Wednesday at 2pm without telling a stranger which specialist.

Financial stuff gets the same treatment. “Money: lawyer call” is fine. “Money: bankruptcy attorney consultation” is a different sentence in a different room. Same for divorce filings, debt consolidation, the job interview you’re taking while still employed, the appointment with the IVF clinic, anything involving a number or a status you’d want to keep close. Beyoncé doesn’t put her vault tracks on Spotify. You don’t need to either.

Other people’s names: be careful. The kid’s friend’s mom doesn’t need to be on your calendar by name. “Family: pickup at park” works. If you’re in a custody situation, talk to your lawyer about what to log and how. The calendar that helps you remember can also be entered into evidence by either side.

The principle is the one Olivia Pope would give you. Only log what you’d be willing to read aloud in a deposition. Specificity is the goal for most things. It’s the wrong move for these.

One honest caution

The labor falls unevenly. People who already feel like they over-track are now being told to track more. If you’re already taxed by this kind of thing, scale it down. Three categories instead of five. Outcome notes only on the big stuff. The point is to make future-you’s job easier, not to add another performance to your week.

Where the AI comes in

This is the Hermione-coded part. Once your calendar has a few months of decent entries, you can ask an AI assistant to do the synthesis. Most of the big ones can connect to your calendar now, with your permission. The two worth knowing about, because most people already have access to one or the other through work or a Google account, are Microsoft Copilot in Outlook and Gemini in Google Calendar.

A prompt that works in either tool, as a starting point:

Look at my calendar between [date] and [date]. Group all events by the category prefix in the title (Work, Health, Money, Family, Home). For each category, list the events with dates, count the hours, and summarize what I spent time on in two or three sentences.

Before you trust it for anything important, run a test on one month. Did it find your prefixes? Did it count right? Did it pick up the ghost entries you backfilled, or skip past them? Make adjustments and see if it works better the following month. If it is still struggling, go to a different system you trust. Better to find out now than the night before an important deadline.

If you live in Outlook (Microsoft 365 Copilot)

Two versions of “Copilot in Outlook” exist, and the difference matters.

The basic version (drafting replies, summarizing a long thread, asking “what’s on my calendar next week”) works with any Microsoft 365 Copilot license. For the system in this post, the relevant feature is that Copilot can read directly against your actual Outlook calendar. The prefixes and color categories you set are the structure it queries. You don’t have to export anything.

The bigger features rolled out in May 2026 are different. Copilot can proactively reschedule 1:1 conflicts, respond to invites based on preferences you set, block focus time, rebook meeting rooms, and cluster meetings onto specific days so you keep uninterrupted blocks. One that pairs especially well with this post is calendar instructions: you give Copilot a written rule for how to handle things on your behalf. For example: “If anyone proposes a meeting before 9am, decline politely and suggest a time after 10,” or “For any event longer than 60 minutes, add a 15-minute buffer after.” These rules sit in your Outlook settings and run automatically.

The honest caveat: these agentic features are not automatically on for everyone with Outlook. They require a Microsoft 365 Copilot license and enrollment in Microsoft’s Frontier program, which is the early-access track for experimental Copilot features. If you’re on a work account, your IT admin has to opt the tenant into Frontier and grant you user access. If you have a personal Copilot license, you opt in yourself. They are also Windows and web only right now, not the Mac desktop Outlook client. If you open your settings and don’t see Calendar Instructions under the Copilot menu, that’s the reason. Ask your admin, or check if you’re on a license tier that includes Frontier access.

If you live in Google Calendar (Gemini)

Gemini’s Google Calendar integration does the same core same-ecosystem trick on the Google side. You can ask it to find, create, and edit events directly, including on shared and secondary calendars (useful if you keep a separate calendar for the ghost entries, and you might want to). The features worth knowing about: better meeting-time suggestions across time zones and working hours, a Daily Brief that pulls priorities from Gmail and Calendar into one view, and an Add to Calendar button inside Gmail that lets Gemini turn confirmation emails (doctor’s appointments, school pickup changes, contractor estimates) into properly titled calendar entries automatically. That last one is the smoothest path I’ve found for the ghost-entry habit. Forwarded confirmations become structured entries without you typing them.

Which one if you have both

Most people I know run work in one ecosystem and personal in the other. You can link Outlook to Google Calendar and vice versa, but it doesn’t solve the problem this post is trying to address. If you want to know why, drop a “?” in the comments. If ten people ask, I’ll write a quick post about it.

In the meantime, the path of least resistance is to run each AI on its native calendar. Your work assistant doesn’t see your therapy appointments. Your personal one doesn’t see the project codenames. Set up the prefixes and colors separately in each. Slightly more setup, no cross-platform glitches.

You can also ask narrower questions. How many times did I take the kids to the doctor this year, and on which dates is a tax-deduction conversation. Which weeks did I work more than fifty hours is a boundaries conversation. What did we spend on the house in 2026 is a should-we-stay conversation.

A real word of caution: the AI will also lie to you with a straight face. This is called hallucinating, and it is a feature of how these models generate language, not a bug specific to one tool. Copilot and Gemini can both miscount hours, invent a meeting that never happened, attribute an event to the wrong person, or smooth a thin week into a plausible-sounding paragraph that didn’t actually happen. Reading directly from your calendar lowers the risk versus pasting raw text into a chatbot, but it doesn’t eliminate it.

So treat the AI summary as a first draft, not a verdict. Skim it against the actual calendar before you paste any of it into a self-review, a tax document, or a conversation that matters. If the ghost entries aren’t there, the labor isn’t there. If the ghost entries are there but the summary doesn’t match them, the AI is the one that needs correcting. The system is only as honest as the human checking it.

The actual point

Self-reviews, year-end reflections, the talk with your accountant, the talk with your partner about whether this year felt fair. None of that is going to become fun. That isn’t the offer. The offer is that the work you already did this year doesn’t have to evaporate between January and the moment someone asks you what you’ve been doing with your time.

Open your calendar settings tonight. Make five categories. Rename the next ten events so future-you can read them. Ten minutes. The rest builds on itself.

Resources

Microsoft Copilot in Outlook

• Calendar Instructions in Outlook and Copilot (Microsoft Support). The official how-to for writing rules Copilot will follow on your calendar.

• Copilot in Outlook: New agentic experiences for email and calendar (Microsoft Community Hub)

• May 2026 Outlook Updates: Calendar Parity, Team Views, Copilot Insights

Gemini in Google Calendar

• Create & manage your calendar events with Gemini Apps (Gemini Apps Help). The official guide for asking Gemini to create, find, and edit calendar events.

• Better time suggestions for meetings with your colleagues using Gemini in Google Calendar (Workspace Updates)

• Google’s Gemini Daily Brief nears launch ahead of I/O 2026

I should mention that I have written, more than once, about my concerns related to

having my voice appropriated by AI for ends I would never consent to. About the way synthetic audio of Black women’s voices, in particular, gets used in scams, in deepfakes, in the small daily violences the internet has made cheap. I am, by my own published record, a person who is suspicious of the thing I just built.

I loved what I made. I also closed the laptop and felt a little sick.

Two things in the same breath. That is the place I was sitting when I read Pope Leo XIV’s Magnifica Humanitas.

I want to say first what the document does. This is not a throwaway. Leo XIV chose his name to put himself in a line with Leo XIII, who published Rerum Novarum on May 15, 1891. The new encyclical is dated May 15, 2026, the 135th anniversary, which is to say Leo XIV is locating his AI document inside the longest-running serious conversation the Catholic Church has had about how technology rearranges power. He is taking AI as a spiritual question, not only an ethical or technical one. He names the technocratic paradigm, the tendency to let “the logic of efficiency, control and profit alone shape personal, social and economic decisions” (paragraph 92). He names what most policy documents will not say out loud, that the speed of AI development is “armed competition” (paragraph 110), driven by geopolitical and commercial dominance, and that calling for ethics-on-top is not enough if the moral framework gets written by the same people who control the data centers.

Two paragraphs really stand out. Paragraph 173 is where the encyclical names the hidden labor: the millions of people, “predominantly women,” doing data labeling and content moderation under brutal conditions, often exposed to the worst of what the internet produces, and the children and adolescents mining the rare earths so the chips can be made. Paragraph 178 names the next colonialism, with data as the new rare earth, extracted from regions of “structural fragility” without their consent, used to train models that will not serve them. These are not gestures.

This is a Pope, on the record, saying that the systems many of us use to write, to teach, to schedule, to build avatars, rest on a chain of exploitation the Catholic conscience cannot pretend not to see.

That mattered to me. I want to be clear about that before I say what I think is missing.

Where the document comes closest to my actual work is its chapter on schools. The Pope writes, in paragraph 140, that “schools are not called to follow the pace of the digital world, but to offer that which the digital sphere by itself cannot provide, namely a shared time for learning and developing trustworthy relationships.”

Yes. Exactly.

He calls for an educational alliance among policy-makers, institutions, and families (paragraph 142), and warns that the asymmetry between business models that monetize attention and parents trying to raise children cannot be solved by telling families to try harder.

I have spent fifteen years thinking about how people learn, and that paragraph names what schools are for. The diagnosis is right. The thing that is being asked of schools is also right.

And.

Here is what the document does not quite hold.

The students sitting in front of a faculty member I support next semester will graduate into a labor market that has already laid off the analyst, the paralegal, the copywriter, the customer service rep, the junior developer, the graphic designer, the translator. Not because those people did anything wrong, but because their employer bought a tool six months earlier than they expected to need it and decided it was cheaper to keep three people than seven. This is the last two years of headlines. The cost of “not following the pace” is being paid by individuals, in real time, with their paychecks.

So when the Pope tells schools not to chase the pace, I find myself wanting to ask: which pace? The pace of the model release cycle, no, do not chase that, that way lies a workshop on a tool that will be deprecated by the time the syllabus is approved.

But the pace at which AI fluency is becoming a hiring criterion? Students cannot afford to opt out of that race, and neither can the adults in continuing education, and neither can the adjunct teaching the AI literacy unit on top of a 4/4 load with no release time. The encyclical’s response to job displacement (paragraphs 151 through 156) is correct on the principles. It calls on the State and on international cooperation to absorb the cost of transition. It is also being read in a context where the State is not absorbing it, and employers are not absorbing it, and so the cost lands on the person who got the email about restructuring on a Tuesday afternoon.

This is the pressure the document names but does not, I think, quite sit with: the people it is talking to are not standing outside the system. We are inside it. I am inside it. I built the avatar. I will probably use it for something. I am also the person who has read paragraph 173 and cannot un-read it, and who lives in a small rural town that data centers are coming to surround. The cooling. The water. The grid. The land. None of this is abstract to me, and almost none of it is something I, individually, can opt out of by not using ChatGPT for a week.

The encyclical’s moral framing keeps the user at arm’s length from the tool. It writes as though the reader is being warned about an environment they might choose to enter.

Most of the document’s actual readers, including most Catholics with consciences who cared enough to read a 245-paragraph encyclical, are already inside the environment. We are using these systems for work we cannot afford to stop doing. We are teaching with them, in some cases, because the next class of students will be hired by people who expect them to know how. We feel both the lift the tool gives us and the weight of what it costs. The document does not yet have a category for that person.

I am not asking for permission. I am asking for honesty about where the addressee is standing.

If we are inside the system the encyclical is warning us about, we are not the first to be.

The encyclical names Martin Luther King Jr. exactly once, in paragraph 124, in a list of “individuals who truly took the dignity of everyone seriously.” It is the kind of mention King gets in most papal documents: lifted up as a moral exemplar, briefly, in the abstract. Inspiration without analysis. Icon, not argument.

But King had this argument. The actual one. Decades before Magnifica Humanitas named the technocratic paradigm, King was naming it, from a Riverside Church pulpit and from inside the long discipline of the Black Church.

In Strength to Love (1963), King wrote: “Our scientific power has outrun our spiritual power. We have guided missiles and misguided men.” That sentence does the same work the encyclical’s chapter on the technocratic paradigm does, in eighteen syllables.

In “Beyond Vietnam,” delivered at Riverside Church on April 4, 1967, exactly one year before he was killed, King said: “When machines and computers, profit motives and property rights are considered more important than people, the giant triplets of racism, extreme materialism, and militarism are incapable of being conquered.” Read that next to paragraphs 92 through 96 of the encyclical. They are saying the same thing, in nearly the same words, 59 years apart. The encyclical does not cite King here. It should.

And then there is Memphis. King was killed in April 1968 because he had gone to support the sanitation workers’ strike. Black men carrying placards that read I AM A MAN. The encyclical’s chapter on the dignity of work is correct. King was killed for the standpoint that produced it. The papal critique and the Black prophetic critique converge on the diagnosis. They do not converge on the cost paid to make it.

I want to name a tension here, because I am a Black Catholic and I cannot pretend it isn’t in the room. King’s theology was forged in the Black Baptist tradition, in a specific Southern Black experience of struggle. It does not sit neatly under a Roman frame. There is something uncomfortable about citing him inside a piece about a papal encyclical, as though the Magisterium gets to fold his prophetic voice into its own lineage and call it ecumenism. I notice that discomfort. I am writing through it on purpose, because the alternative is to let the document mention King in paragraph 124 and let no one ask why his actual argument is not the spine of it. The two traditions are not interchangeable. Placing them side by side is not synthesizing them; it is refusing to let one stand in for the other.

If the encyclical wants the dignity tradition King died for, it has to mean more than putting his name in a list.

There is one more thing the encyclical, and most of the AI discourse it joins, keeps doing that I want to name.

It treats AI as the defining problem of our moment. Not a problem. THE problem. Its 245 paragraphs are organized around it. The “new things” of our era are, almost by definition, AI things. Even the chapters on work, war, and education orbit it.

When we frame every problem as an “AI problem,” we let the older structures off the hook.

The encyclical names data colonialism (paragraph 178) and labels it new. But extractive capitalism did not start with data. It started in the silver mines of Potosí and the cotton fields of Mississippi. The encyclical names automation’s labor displacement (paragraph 150) and treats it as part of “the fourth industrial revolution.” But automation has been displacing workers since the first industrial revolution. Surveillance capitalism predates ChatGPT by a quarter century. Misinformation predates deepfakes by all of history. The cobalt mines in the DRC were already exploited before they were feeding chip fabs.

This matters because the response we organize matters. If “AI” is THE problem, we get an AI regulation regime, possibly a good one. We may even get a Catholic alliance around AI ethics. And we will, six months later, discover that the underlying disease (what King called the giant triplets, what Catholic Social Teaching since Rerum Novarum has called the disorder of capital over labor) is still here, just wearing different clothes.

The encyclical, to be fair, does sometimes gesture at this. Locating itself in the Rerum Novarum tradition is itself an act of saying “we have been here before.” But its center of gravity, its 245 paragraphs, are organized around AI as the defining “res novae.” Its imagination is AI. Its target is AI. The response it convenes will be an AI response. The framing does work. The framing also does damage, because it can let everyone in the room agree to regulate the new thing rather than confront the old thing.

Instructional design has been running a version of this panic cycle for as long as I have been in the field. Calculators. The internet. Wikipedia. MOOCs. ChatGPT. Each new technology gets framed as the existential threat to learning. Each time, the underlying structural questions (who gets supported, who has time to study, whose knowledge is treated as legitimate, who pays for the tools) stayed in place. The panic served as a heat sink. It absorbed the attention that might have gone to the older, slower, harder argument. By the time we recovered, the next technology was already arriving.

AI is the latest panic. It is also a real thing, with real costs, and I am not arguing that we should ignore it. I am arguing that if it becomes the only frame, we will regulate AI and discover that poverty is still here. Automation is still here. The disorder King and Leo XIII both named is still here. Wearing what comes next. And we will still be standing inside it.

The encyclical closes with the Magnificat, Mary’s song of reversal. The proud scattered. The mighty cast down. The lowly lifted up. The hungry filled. The rich sent away empty.

If we are going to take this document seriously, the song has to reach further than the document quite admits. The Pope points correctly at the developers, at the technocratic mindset, at the lack of regulation. But the institutions calling for the response, the Church, the university, the employer, the state, have not yet absorbed the cost they are asking individuals to absorb. There is no parish-level AI formation am aware of (please correct me if I am wrong). There is no diocesan retraining of priests for the questions their parishioners are actually asking. Most universities are running AI literacy on the unpaid time of faculty who are figuring it out on the side. Most workplaces are passing the cost of fluency through to the employee, whose next performance review will quietly reflect whether they kept up.

The Magnificat is not a poem about somebody else’s reversal. It is a song about the order of the world being turned over, including the parts of the order that the singer benefits from. If the Church and the institutions that share its concern want this document to be more than a beautiful diagnosis, the lift will have to land on their budgets, their staffing, their formation programs, and yes, eventually, their relationships with the data centers being built outside small towns like mine.

I will keep using the tool I just built, with my eyes open about what it cost to make and what it will cost to run, and the other equally important ethical considerations.

I know that is not a clean place to conclude. The encyclical does not promise a clean place. It promises, if I am reading it right, that the Lord has already shown the strength of his arm, and that the song is being sung whether we are ready for it or not.

Maybe the most honest thing I can do, this week, is re-read paragraph 173 out loud.

It’s Wednesday. 11:43pm. The email starts with “Hi Dr. Horton, I’m so sorry to bother you this late, but…” and I already know what the next paragraph is going to say. I have been on the receiving side of an email like this enough times in my own life to know how the sentence ends.

I have a late policy. It is in my syllabus. It is in the section the office for the Vice Provost for Faculty and Academic Affairs helped me write. My institution gives me good guidance on accommodations, on dean of students referrals, on what counts as documented and what doesn’t. The routine cases are covered. Most of them never reach my inbox at all.

This is not a routine case. This is the call.

The case the policy didn’t write

I used to watch a lot of court TV. (Don’t judge. Or do. Fitting.) Judge Mathis is up there making twenty rulings (I exaggerate; it’s two…everyone knows that!) before lunch and you almost never see him improvise. Not because he’s a robot. Because he has a bench. He has a clerk. He has the docket in front of him, the statute on the wall behind him, and a rule he named out loud before anyone walked through the doors. Judge Judy will tell you to sit down before you have even finished your sentence. The structure is doing half the work.

A faculty member with a syllabus has a version of that structure. The syllabus paragraph is the closest thing I get to a codified statute, and for most of what comes across my desk, it is enough. The student missed the deadline by a day, took the standard penalty, moved on. No call to make. The policy already made it.

But every term, every single term, there are the cases the syllabus did not write. The student whose grandmother died and whose paperwork is still in the registrar’s queue. The student whose ADHD documentation says one thing and whose week has been another. The student whose laptop did, in fact, die, and whose week-six pattern is starting to look like something other than carelessness. None of those are covered by the statute. All of them are now mine.

That band of decisions, the discretionary zone, is where bias has the most room to work. The routine cases do not need me. The cases that need me are the ones where I have room to move. And I have been making those calls for years with no scaffolding around them at all.

What I wanted, and what we built

So I built a thing. More honestly, a few of us built a thing together, and I will speak for the part of it I cared about most.

I wanted a chambers. A small, private, ten-minute room I could step into before I responded to the 11pm email. Not to make the decision for me. To make sure that whatever I decided at 11:43pm on Wednesday looked like what I decided at 2:15pm on Tuesday and at 9am on Friday. Not “consistent” in the way a robot is consistent. Consistent in the way a judge who has been doing this for thirty years is consistent. Because she decided some things in advance, on purpose, when nobody was waiting.

The tool runs in two modes.

Before the term starts, I sit with a design-mode questionnaire and answer a handful of questions about what kind of work this is, what comes after it, what I want the policy to teach, how much grading bandwidth I actually have, and what my institution requires. (That third one was the question almost nobody had answered when we first started asking. Most faculty had a policy. Very few had picked the lesson.) From my answers, the tool generates five things at once. Edit any answer, and all five regenerate together.

1. A written late policy with its rationale visible. Not the syllabus paragraph, a short document that explains, to me, why the policy is what it is. I keep this one for myself.

2. Syllabus-ready language in three tones (warm, formal, direct). I pick one and paste it into the syllabus.

3. A five-branch decision tree, exportable. Documented accommodation, grace window, token, make-up, escalation. The tree is what gets walked at midnight. The syllabus paragraph is not.

   

4. A response template per branch. When a request comes in, the tool drops the right template into an editable box. I edit. I copy. I paste into my own email. The tool never sends.

5. A decision log, categories only, no PII. For self-audit at midterm and at the end of the term. Was I consistent? Did my policy land where I expected?

Mid-semester, when the email comes in, I step out of the inbox for sixty seconds. I categorize the situation. Within the grace window or past it. Illness, family, tech, time, or undisclosed. Tokens used so far. The tool walks the situation through the tree I built in week zero, drops the right template into a draft, and waits for me to do the actual sending.

That is the whole product. It is meaningfully smaller than it sounds.

What I deliberately didn’t build

The list of things this tool does not do is longer than the list of things it does, and the omissions are the design.

• It does not send email. The faculty member’s voice and judgment belong to them. The tool is upstream of the send button, never past it.

• It does not connect to the LMS or the gradebook. The moment we pull a roster or a deadline, the tool becomes an LMS tool, with all the audit trails and vendor agreements that implies. It is FERPA-safe because it never sees the student.

• It does not score the student’s reason. A category is just a label for the faculty member’s own log. “Illness” is not adjudicated. “Undisclosed” is not penalized. The reason is not the input, the timing is.

• It does not auto-extend, auto-escalate, or auto-anything. Every branch ends in a faculty member’s decision. The tree is a default, not a verdict.

• It does not see student-level patterns. That’s a feature, not an oversight. The argument is that the loop where the same student keeps asking is supposed to be broken by conversation, not by a counter… because a counter flattens “going through something” and “gaming the policy” into the same number.

This is the cost of the architecture. The tool will not “save you time on late work” in the way that a sender-and-syncer would. It will instead make sure that when you do spend the sixty seconds, you spend them on the part that actually requires a human. It will also help you reflect on your policies with data you don’t normally have access to.

The docket

There is a log. I want to talk about it specifically, because it is the most data-heavy thing in the product and the shape of it matters.

What gets stored, per decision: the date, the week of the term, an assignment code, a category, the branch I landed on, and the action I took. That is the entire row.

What does not get stored, ever: the student’s name. Their email. The wording of what they wrote. My reply. The grade. Anything that could connect a row back to a person. If I shared my log with a colleague teaching the same course tomorrow, no student in either of our classes would be re-identifiable from it.

What does not get stored, ever:

• the student’s name,

• email address,

• the wording of what they wrote,

• your reply, the assignment grade, or

• any identifier that could connect a log entry back to a person.

The log exists to answer exactly one question, and I want to be precise about which one. It is not “who deserved what.” It is “did I rule the same way on the same kind of case.” It is a record of my own behavior. It is not a record of theirs.

I think about Hot Bench, where three judges argue with each other before rendering a verdict. I do not have two colleagues sitting next to me at midnight. The log is the closest thing I get. At midterm, I look at it and I ask: did I grant every illness request but treat the undisclosed ones harder? Are my extensions clustering in weeks six and seven, which means the course is pacing wrong, not that the students are weak? Have I been escalating more this term than last? Why?

The log cannot tell me what to do about any of that. It can only tell me what I have, in fact, been doing.

The one step I cannot skip

Before the tool will generate a policy, it makes me answer three questions I cannot click past. Who in this course typically asks for extensions. Who never does. Whether there are students whose late work does not look like late work, but is. Whether the policy I am about to seal gives me cover to say no to something I would have said yes to last year.

I did not include these prompts because I think a five-minute reflection cures bias. I included them because I have watched what happens without them. A faculty member who has historically extended more easily for the students who remind her of herself will, with no pause and no audit, build a tree that does exactly that, quietly, and now with documentation. The Prompt is the one place in the tool that interrupts that. It is the one step I made non-skippable.

A structured policy is a tool, not a moral position. If it lets me say no to things I would have said yes to before, that is a flag. It is not a feature.

Three risks I am not pretending away

A tool like this can become cover for refusing accommodations that should have been granted. If I start saying no to things I used to say yes to because the tree said so, the tree is failing me. I am watching for it.

Response templates can feel mechanical. I tried to write them with warmth. Warmth is fragile in a template. I edit every one before I send. So should you.

The log will tempt some of us to compare students. Don’t. It is not what the log is for. The only question it can answer is whether I have been consistent with myself.

What’s next

V1 ships with the five outputs above. V2 will add a midterm dashboard that surfaces patterns I will not otherwise see (”you granted twelve extensions this semester, mostly in weeks six and seven, what was happening in your course at week six?”), and a small library of policies other faculty in similar courses have written.

The things that will never be on the roadmap are the same things I refused to build into V1. Auto-emailing students. LMS integration. Scoring the reason. Those are not unimplemented features. They are the line that lets the rest of the tool exist.

If you teach, or if you run a teaching center and you have been asked the late-policy question one too many times, the tool is open. It works in a browser. It requires no account. It stores nothing on a server. Spend thirty minutes with it before your next syllabus. Spend ninety seconds with it the next time a 10pm email shows up.

I hope you’ll feel the difference at week fifteen.

I want to tell you about them, and I want to do it in one post instead of fifteen, because what I’m trying to find out is bigger than any single tool. I want to know whether there’s appetite for two things: a teach-along, where I walk through how I actually made these, and a critique-along, where we pull each one apart and I tell you what it cost, what I’d cut, and what I’m still not sure about.

I’ll get to the asks at the bottom. First, the apps.

They sort, weirdly cleanly, into three groups based on what they refuse to do.

Slow down before you decide

Four of the things I built are upstream tools. They do not save you time. They make you spend twenty more minutes thinking before you write a sentence, send an email, or post a rubric. They are friction, on purpose.

Late Policy is a tool I almost didn’t build, because the obvious version (read the student’s email, draft the response) is exactly the version no faculty member should ever use. The tool I made instead asks one question first: what is your late policy for? Until you’ve answered that, the tool will not generate anything. There’s a Fairness Prompt you cannot skip. There’s a self-audit log so future you can read past you and notice the drift. It is, to be clear, a slower tool than the obvious version. That is the entire point.

The Workbook came out of a panel I moderated on how educators are learning to use AI in their work. The panelists disagreed, productively, and I went home, transcribed the conversation by hand, and noticed I had filled half my margin with my own questions. So I built a workbook. Sixty to seventy-five minutes, self-paced, with six real tensions on a continuum and a prompt asking where you stand and why. The workbook refuses to take a side on AI in teaching. The temptation, when you build a thing for educators about a contested topic, is to give them a handrail. We held off. Your marks on six lines are the handrail.

Reflection is the thing I most wish I’d had six years ago. Twenty minutes per course, one document, five prompts designed to be hard to ruminate inside (what worked, what broke, what surprised you, one question, one decision). While you write, the tool flags hedges in real time (the “didn’t land,” the “seemed checked out,” the “kind of”) and tags each with a one-line concreteness fix. The paragraph then scores against five qualities from the reflective-practice literature: specific, multi-lens, evidence-grounded, asks why, closes the loop. A separate shelf tracks one document per course across the term, opened every July when you plan fall. Around the per-course reflections, four surrounding artifacts point at promotion and tenure: a response-to-feedback memo, a philosophy addendum, a per-cycle retrospective ritual, and a trajectory view across terms. For the one or two courses a year where you’re running something you’d want to defend to skeptical colleagues, a scholarship toggle wraps the reflection in research scaffolding (guiding question, study design, IRB posture, target venue, and a falsifier named in advance). Year over year, the folder becomes the most useful single piece of evidence about your own teaching that you own.

TILT Studio started because every assignment ships with two rubrics, and the one in your students’ heads is usually the one they’re writing for. The studio takes a rubric you already use, runs a ten-item transparency diagnostic on it, and walks you through making the implicit explicit. It aligns each criterion to a learning objective and a Bloom’s level so the rubric measures what you’re actually trying to teach. The Winkelmes studies put the effect at around twelve percentage points, concentrated in first-generation, low-income, and underrepresented students. The hidden rubric is not neutral. It’s a tax. The studio is built to make it optional.

The thread through all four: none of these tools sees a student. They work on you.

Narrower tools, on purpose

The next three are tools that hand back time without ingesting student data. Every obvious version of each one is a FERPA hazard, and the interesting design move was building the smaller version that does less than the technology can plausibly do.

The Accommodation Implementation Handbook never asks for the letter. It asks for two things from the faculty member: the accommodation category, picked from a standard list, and the activity composition of the specific course. From those two inputs, it flags the activities where the accommodation is genuinely standard, and flags more loudly the ones where it’s easy to forget, like the wet labs and the field days where the LMS-shaped mental model doesn’t match the activity at all. It drafts a first message in three registers that you adapt rather than send. The handbook surfaces the two thirds of an accommodation that the letter writer never imagined.

Recommend came out of the realization that the hard part of writing a letter has never been the writing. The student remembers their projects, their failures, the moment the lab finding turned out to be a prompt-format artifact. The writer remembers an impression. The student wizard structures evidence on the student side. The writer side is a three-pane workspace where every sentence in the draft cites its source. Hover any claim to see which piece of evidence it came from. If the AI invented something, the citation is empty and you can see it at a glance. The tool doesn’t write better letters. It just stops throwing the student’s evidence away. That turns out to be enough.

Hearth is the one I’m personally most fond of. Six small instruments for the working teacher: Misconception Predictor, Analogy Generator, Worked-Example Builder, Discussion Prompts, Question Auditor, Case-Study Generator. You bring a piece of material from your course (a topic, a reading, a question you wrote at 1 a.m. on Tuesday), and the tool brings a pedagogical lens. No rosters, no grade book, no analytics, no predictive flags about which students are “at risk.” Faculty kept asking for tools that did not watch their students. So I built those.

The thread: every one of these was a louder, more powerful product before I narrowed it. The narrow version is the one worth shipping.

The friction is the feature

The last eight are student-facing, or at least student-adjacent, and they all share an unfashionable bet: the value of an educational tool in 2026 is not its ability to be helpful but its ability to be well-shaped resistance.

WHEREFORE is a civic-reasoning prototype built around a mascot named Quill, whose defining feature is that he will not tell you the answer. Not when you ask politely, not when you ask three times, not when you tell him you’re stuck. Phase four has a premise bank where half the items are blurred under a frosted overlay, and tapping one to reveal it costs an “Ink,” a soft currency tracked at the bottom of the screen. The point isn’t to prevent peeking. The point is to make peeking visible, to yourself and to your teacher. The teacher has a toggle that flips the same URL into a feedback view with three tags: PRAISE, FLAG, REVISE. Three was the right number; more granular tags died on contact with actual teachers.

SYLLABUS://OVERRIDE is a fifteen-minute browser-based co-op game where students play cyberpunk operatives infiltrating an “Academy mainframe” that is, every time, the course syllabus. Three timed protocols and two interludes are stitched from a real syllabus, with every wrong answer pulled from a real misconception that has shown up in office hours. A quiz tests recall; a game tests recall under load, which is how students will use the information at 4:55 PM when a deadline shifts. The Uplink Triage interlude routes distress signals to actual campus support resources (CAPS, the food pantry, emergency funds, 988), so when a student hits week ten with a bounced rent check, “Emergency Funds” is already in their head. Accessibility is the default, not the setting: dyslexia font, larger text, high contrast, extended timers, and keyboard navigation on by default; sync-burst effects off.

Voice Agent is the version of the same bet, scaled up, and so far the most popular thing I’ve shared with faculty. The use case people keep returning to: feed the agent your syllabus and let it answer the dozen FAQ-shaped questions students send you by email every week (when is the midterm, what’s the late policy, where do I submit), with guardrails. By the last module of the configurator, an instructor has picked a friction level on a five-point dial, written a system prompt their class actually defines, wired tools the agent can call mid-call (book office hours, escalate to a human if a student calls in distress), and run live test scenarios against a real model. The course is opinionated. It won’t let you ship with a fifty-character system prompt or upload anything identifiable about a student.

PromptCraft is the seven-minute warmup that sits before Room 204. You pick a real classroom problem from three options that name the things actually eating your week, then write a prompt that works by reading three drafts side-by-side and pulling apart what makes the strong one strong. The output is a recipe you can copy into Claude or any other chat on Monday morning and watch it build. No login, no setup, no code. The bet is that the fastest way to teach prompt craft is to put three versions of the same prompt next to each other and ask the learner which one will actually work.

Room 204 is a small game I made to teach teachers how to build AI tools. You play a teacher in a fictional classroom. Monday: you notice the gap. Tuesday: you write a prompt, picking a Socratic Tutor or an Exit Ticket Bot (only two archetypes, on purpose) and choosing rules from a list where half are subtly bad. Wednesday: class day, and the prompt fails in three different ways for three different fictional students. Thursday: you fix it. Friday: you ship it, and the chatbot in the live-demo tab is actually running on the prompt you just wrote. Fifteen minutes, no login. The thing you build does the thing.

Git Playground renders Git’s invisible geometry as three colored panels (working directory, staging, committed); files physically move between them as the student types; and a simulated teammate named Riley can be summoned to create a merge conflict on purpose. A merge conflict isn’t an error. It’s git asking you a question. You answer it by editing the file. Rebasing, cherry-picking, the reflog, force-push, and amend are deliberately left out, because the failure mode of Git education is throwing every feature at students before they have the model.

From Blog Post to Animated Video is an eight-scene workshop guide for turning a Substack post into a thirty-second animated video using Runway, ElevenLabs, and CapCut. I built it because I made one, and people started asking how. The teaching insight: the brief is the lesson. The first twenty-five minutes of the workshop are deliberately tool-free, because the strongest predictor of a bad final video is opening Runway before the brief is written. You are writing a script document, not a prose document. If a section reads like a paragraph, it’s wrong.

How to Make a Digital Story is the longest thing I’ve made, a twelve-step field guide for faculty teaching digital storytelling and the students who bring the work to life. Shape before sentences. Draw badly on purpose. Record thirty seconds of room tone at every session. Score the silences, not the sentences. Build the skeleton (radio cut) before the skin (rough, fine, picture lock), and name the passes in your assignment deadlines or students cannot diagnose a muddy cut. Five tool stacks, a scrubbable example, a drag-and-drop storyboard, six downloadable templates you can copy-paste into your syllabus.

The thread through all eight: the work of doing the thing is the thing being taught.

What this was actually about

Building fifteen of these in six weeks isn’t normal. A couple of these looks I’d send down the runway again tomorrow. A couple I’d cut from the collection before the judges saw them. The truthful version is that I’ve been thinking about teaching and AI for a long time, the tools available right now are mostly the wrong shape, and once I figured out I could make the smaller, better-shaped version, I couldn’t stop until I’d drafted them.

I learned things I want to teach. I also have a running list, per tool, of what each one cost. Hours, dollars, the back-end apps I had to learn to use just to make these run, the specific tradeoffs I made and the ones I’m still not sure about, the design decisions that landed in user testing and the ones that quietly fell over.

What I want from you

Two questions, in order of how much I care about the answers.

First: would you want to learn how to make these? Not the philosophy. The build. I’m thinking about a teach-along, three or four sessions, where I walk through the actual workflow I used (the prompts, the iteration, the moments I scrapped a whole approach and started over) and you build a smaller version of one of these for your own course or department by the end. If that lands for you, tell me in the comments. I’ll size the demand before I commit to a schedule.

Second: would you want to read the critique series? A post per tool, with the cost breakdown, the things I would cut if I rebuilt it, and the specific places I think I got it wrong. I’d write these with help from two or three people who have used the tool in their actual teaching, so it’s not just me grading my own homework. If that lands for you, tell me in the comments.

Third, less an ask than an invitation: tell me in the comments which of the fifteen you want me to dig into first. I’ve been hovering between three of them and I’d rather you broke the tie.

And if you teach, and you have a use case for one of these that I should be thinking about, just hit reply. The whole reason there are fifteen of these instead of one is that I kept finding faculty problems I hadn’t thought to look for. Tell me yours.

That’s the runway. Fifteen looks, six weeks, no commercial breaks. None of them is the final shape. All of them are easier to think about with the judges in the room.

I was on a panel last Saturday at AIxED, an AI in education conference in Boston. The panel was called Real Educators Learning AI in Real Time, and the framing came from Sideby: how educators are figuring out how to use AI in their actual work, not what AI can do. Kippy Smith moderated. The other panelists were Yousuf Marvi, an 8th-grade math teacher in Irvine Unified, currently in his first year teaching English Language Development, and Anne Fensie, who directs the Center for Teaching and Learning at the University of Maine at Presque Isle and works mostly with adult learners in a competency-based program.

Kippy opened by asking each of us for our why for using AI. Anne named learning itself. She wants AI to help where help is appropriate and not to flatten the productive struggle that builds the skill. I named educator agency, with a debt to Punya Mishra: no technology was originally designed for education, and educators have always had to adapt what was built for somebody else. AI’s offer is that it puts educators in the designer’s seat for once. Yousuf named his students. The multimodality of current AI means he no longer has to rely on text alone to understand what a novice language learner is doing in geometry. Kippy named the pattern: none of us led with personal efficiency.

The conversation kept circling back to friction. Yousuf put it most cleanly with an effort-to-output ratio. When the ratio is close to one, learning is happening. When it drops too low, learning collapses. Anne quoted Daniel Willingham (memory is the residue of thought) and offered a triumph and a disaster from her own practice. The triumph: she learned enough Python through AI assistance to scrape 30,000 email addresses for her dissertation survey, something that had previously been out of reach. The disaster: she followed AI’s instructions during a database update and overwrote two-thirds of her course content with no backup, then spent ten hours of a conference manually restoring it from change history. The disaster made her more critical of every output. I argued that real learning shows up in AI’s failures, where students see why human expertise matters, and that my work with faculty is mostly about preserving the friction that AI is engineered to remove.

Two audience contributions reshaped the room. Stephen, who had spent the prior week at an agentic AI corporate event, pushed back on schools designing learning from the inside out, starting with learning instead of with outcomes. His line: curriculum should be designed to uncover content, not cover it. Jenna, an AI product manager who sits on a school council where her district is barely engaging with AI, asked the practical question: how do you actually start. Anne’s answer was small daily reps. Mine was to stop believing the hype that everyone else is ahead, and to start with real low-stakes problems. (Mine: a tool that reads my refrigerator and suggests dinners that work for both my husband’s allergies and my weight goals.) Yousuf’s was the sharpest reframe of the panel. If you cannot make the case for AI in your classroom without using the words ChatGPT or Gemini, the work is on the AI literacy side, not on the resistant-teacher side.

The closing frame the room landed on came from another audience contribution: the problem of practice. AI literacy without a real problem to solve is technology for its own sake. With a problem of practice in hand, AI becomes a possible role in a real solution, not the solution itself.

There is a seventh I couldn’t get in there in time. It is the one I am still sitting with, and the one I want to give more room than a prompt.

It is agentic AI.

I should start by saying what I love about it, because the version of this conversation where I open with my worries doesn’t sound like me, and it isn’t true to my experience.

I love that agentic AI puts non-programmers in the design seat. I have built a handful of tools in the last year, some for productivity, some for reflective practice. The fact that I could build them at all is the new thing. That access matters. It is also, in a roundabout way, what pushed me toward pursuing more formal programming background. Not because the agents made me feel inadequate, but because they made me feel adjacent to a thing I wanted to understand from the inside.

The design seat is not the driver’s seat

The first reason is the one Kishore Aradhya named at AIxED, and I keep returning to it. Agents are confident when they are wrong. They mislabel. They miscategorize. They invent metadata that looks right because it conforms to a pattern, not because it is true. Non-programmers are less equipped to catch when something looks right but isn’t. I am fortunate to work at a school of computing, where colleagues have been patient and generous in helping me debug when things go off the rails. Not everyone has that safety net.

The second reason is execution. Agentic tools run commands on your machine. The “approve this action” prompt can become theater if you do not fully understand what you are authorizing. Prompt injection is a real concern. I have worked through tutorials with three or more layers of redirects, where the first two were benign and the later ones raised yellow and red flags about where and how my data was being committed. Some folks in our IT department are excited about what I have been building, and they have also been clear that real vetting takes time before any of it touches sensitive data. I think that pacing is right.

The third reason is privacy. Pointing an agent at your folders means handing it whatever lives in them: tax documents, draft manuscripts, FERPA-protected student data, IRB materials. Even with training opt-outs, file contents travel to company servers. Most universities have not issued guidance that keeps pace with the tools, and faculty are largely navigating this on their own.

None of those are dealbreakers. They are problems with solutions, mostly involving time, policy, and someone to call. The harder question is the pedagogy.

Mapping onto a value is not the same as cultivating it

Here is where my concern actually lives. Agentic AI is increasingly being pitched as scaffolding for the cognitive moves we want students to develop themselves: metacognition, self-regulation, inquiry. The marketing language is not subtle.

When OpenAI launched ChatGPT Study Mode last July, the product page described its behaviors as “proactively developing metacognition and self reflection, fostering curiosity, and providing actionable and supportive feedback.” Read that again. The agent develops metacognition. The agent fosters curiosity. The agent provides the feedback. The student is the object of the verbs, not the subject.

Anthropic’s Claude for Education Learning Mode, launched a few months earlier, frames the agent as guiding “students’ reasoning process rather than providing answers,” prompting them with questions like “How would you approach this problem?” and “What evidence supports your conclusion?” The framing is more careful, but the agent is still the one asking the metacognitive questions. The student is responding to them.

Ethan and Lilach Mollick’s widely cited paper on classroom AI prompts puts the design pattern most plainly: “AI Language Models can help students engage in metacognition as an AI coach by helping direct students to engage in a metacognitive process and articulate their thoughts about a past event or plan for the future” (Mollick & Mollick, 2023). A peer-reviewed framework published last year in Frontiers in Education calls AI “a metacognitive partner” and reports that “AI quantitative feedback improves the accuracy of metacognitive monitoring.”

The bundling shows up most cleanly in a September 2025 EDUCAUSE Review essay co-authored by, among others, Paul LeBlanc and George Siemens. They describe the new student skill profile as developing “metacognition (thinking about their thinking) and meta-emotional skills (managing their emotions), as well as the ability to design and refine AI agents.” Metacognition, emotional regulation, and AI-agent design get folded into the same sentence, as if learning to manage your own mind and learning to manage a tool were the same kind of skill.

Mapping onto a value is not the same as cultivating it. Sometimes the closest match is the thing most likely to displace what we hoped to grow.

Four patterns worth naming

The same elision shows up across the field in patterns I think are worth naming out loud, so we can argue with them as patterns instead of arguing with each product one at a time.

The reflective scaffold pattern is the most common. The agent prompts the student to notice their thinking. What’s one thing you’d do differently? Where did you get stuck? My worry is not that those prompts are bad. It is that the noticing was the skill we were trying to build. If the prompt arrives on time, every time, calibrated to the right level, the student never has to develop the internal cue that says this is the moment to stop and check. The scaffolding never comes down because the scaffold became the skill.

The persistent project partner pattern proposes that an agent that remembers your work across sessions is building your follow-through. The pitch is that the agent holds the through-line so you can re-enter the work. Persistence, though, is something a person does, not something an agent provides on a person’s behalf. An agent that remembers your contradictions for you may not be building your persistence so much as building its own continuity while yours quietly atrophies.

The role-player pattern asks the agent to simulate a journal editor, a skeptical colleague, a peer reviewer. The assumption is that an agent with goals and constraints meaningfully approximates the human in that role. That is a sizable assumption. Real editors and colleagues bring institutional context, professional stakes, and accountability. An agent can simulate the shape of those interactions without the weight. The piece I worry about most: students may learn what an editorial exchange looks like without ever building the judgment, the resilience, or the read on power dynamics that real editorial exchanges develop in a person.

There is a fourth pattern that does not have the same problem, and I want to give it real space. The design-the-agent pattern has students making agents instead of using them. Constraining them. Arguing with their behavior. That move treats AI as something to be made legible and contestable rather than simply consumed. The cognitive work stays on the student because the student is the designer.

The burden of proof

None of this is an argument against experimenting. Marc Watkins put a version of this in the Chronicle last fall: higher education, he wrote, “must explore ways of curbing the use of agentic AI tools to automate work that students are supposed to do on their own.” Punya Mishra asked the question more compactly in February: how do we build agents that work with students, not for them?

That is the question I want to ask out loud, alongside both of them. The burden of proof sits with us to show that agentic designs build the agency they claim to support, rather than simulate it convincingly enough that we stop asking.

He died. Again. For the 47th time. Melinoë turned into a red silhouette of herself, spun, and disappeared. Then he was back at the Crossroads with nothing but a handful of bones and the reminder that Time Cannot Be Stopped.

And instead of quitting, my husband leaned forward. Swapped an Arcana Card. Tried a different weapon. Went back in. I was sitting next to him on the couch, not holding a controller, but I couldn’t look away. Not because the game was pretty. Because I’d seen this before.

This is Hades II, Supergiant Games’ roguelike sequel that drops April 14. You play as Melinoë, daughter of Hades, fighting your way through the underworld (or up toward Olympus) to defeat Chronos, Titan of Time. You will die. A lot. The game is counting on it. Every mechanic, every system, every line of dialogue between runs assumes that failure isn’t just possible. It’s the process.

Not in a game review. In a classroom.

I studied this

Years ago, I had the honor of taking serious gaming courses with Dr. Carrie Heeter at Michigan State, where we learned the categories of serious games and how they create sandboxes for students to learn without penalties. The idea wasn’t abstract. We played games, categorized them, pulled apart the design decisions that made risk feel safe. The theoretical spine came from James Paul Gee, who in What Video Games Have to Teach Us About Learning and Literacy (2003) named what he called the “psychosocial moratorium” principle: games create spaces where learners can take risks, fail, and try new identities without real-world consequences. That’s not a metaphor. It’s a design decision.

Gee’s framework has held up. Bacalja et al. brought 21 scholars together in Postdigital Science and Education in 2024 to revisit his learning principles, and the consensus was clear: the core ideas are still doing work two decades later. And a 2025 study in the International Journal of Educational Technology in Higher Education found that university students in competitive sandbox game environments showed stronger critical thinking and collaboration than those in teacher-led instruction.

So when I watched my husband pick up Hades II and start dying his way through Erebus, I wasn’t playing. I was sitting next to him, watching the design work in real time, recognizing every principle I’d studied with someone who helped define the field. He was fighting his way through the underworld. I was diagnosing the learning design.

Four mechanics, four lessons

Let me walk you through the design decisions that matter, because Hades II isn’t just fun. It’s a learning environment that does things most courses don’t.

God Mode is adaptive scaffolding. Turn it on and you start with 20% damage reduction. Die, and it goes up by 2%. Cap is 80%. No achievement penalties. No shame messaging. The game notices you’re struggling and quietly makes things a little easier each time. Faber et al. published a study in BMC Medical Education in 2024 showing that adaptive scaffolding in game-based learning enhanced both performance and engagement by optimizing cognitive load. God Mode is that finding built into a difficulty slider.

Now think about your course. A student fails your midterm. What adapts? Does the material get 2% more supportive? Does anything in the course design acknowledge that this student needs a different entry point? Or do they just get the same exam again next semester?

Arcana Cards are cumulative knowledge. These are permanent upgrades that persist across every run. There are 25 of them, and they have synergies. Early on, you’re just collecting whatever you can unlock. Ten runs later, you’re building optimized combinations because you’ve learned which cards work together. That’s the shift from surface-level recall to expert schema-building, and it happens because the game lets you keep what you earn even when you lose.

Boons are scaffolding in the moment. During a run, gods offer you temporary power-ups. Three choices, each calibrated to your current build. The game is handing you support inside the Zone of Proximal Development without you filling out a form or visiting office hours.

The story keeps going when you die. NPCs have new dialogue after every failed run. Relationships deepen. Plot threads advance. Death isn’t a reset. It’s a chapter break. The narrative tells you, over and over: this attempt counted, even though you didn’t win.

Powers and Moore reviewed failure state mechanics in game-based learning in TechTrends in 2021 and found that designed failure states show real promise for meeting instructional goals. Hades II isn’t stumbling into good pedagogy. It’s executing it.

Now compare that to your midterm

Manu Kapur published Productive Failure: Unlocking Deeper Learning Through the Science of Failing in 2024, and his four mechanisms map almost embarrassingly well onto what this game does. Activation of prior knowledge? That’s every run where you remember what killed you last time. Attention to critical features? That’s the boss pattern you finally recognize on attempt six. Explanation and elaboration? That’s the moment you start choosing Boons not randomly but deliberately, because you can articulate why one synergy works better than another. Organization and assembly? That’s the Arcana build you’ve refined across 30 runs into something that reflects everything you’ve learned.

Kapur’s framework. A roguelike’s execution.

And the macro-level data backs up what the game-level design suggests. Barz et al. published a meta-analysis in Review of Educational Research in 2024 and found a medium effect for overall learning (g = 0.54) and a strong effect for cognitive learning (g = 0.67) in digital game-based environments. Games work. The research is not ambiguous about this.

Meanwhile, in spring 2026, more than a dozen universities have abandoned AI detection tools entirely. Not because the tools failed at catching AI (they did, with false positive rates hitting 61% for non-native English speakers). But because the deeper realization was worse: the assessments themselves were so easy to outsource that the whole integrity apparatus was beside the point. The assessment crisis of 2026 isn’t an AI problem. It’s a design problem. Games figured this out years ago.

So what would you actually do?

I’m not going to pretend this is simple. The sandbox works. The question is whether institutions are willing to create the conditions for faculty to build one. Most instructors aren’t given the time, autonomy, or support to redesign a single assignment, let alone rethink a whole assessment architecture. Course caps, standardized assessments they didn’t choose, adjunct workloads that don’t include prep time for innovation. Willingness isn’t the bottleneck. Capacity is.

But the gap isn’t only structural. It’s also about knowing what tools exist. Delphinium is a Canvas-based gamification framework that lets instructors build choice-driven, leveling-up course structures inside the LMS they already use. Students pick pathways, accumulate progress, and see their growth visually. It’s not Hades II. But it’s the Arcana Card system adapted for a course with 150 students and one overworked instructor. Canvas Mastery Paths, which is already built into Canvas and costs nothing extra, lets you set a source assignment (a quiz, a graded discussion, anything with a score) and then automatically route students to different content based on how they performed. Score below 70, the system assigns additional support material. Score above 85, the student skips ahead to the next challenge. Students don’t all have to see the same content in the same order. That’s God Mode logic inside a tool most faculty don’t even know they have. MyQuest takes it further with quest-based course structures where students progress through narrative arcs with multiple pathways and built-in feedback loops. For Moodle users, Level Up XP adds experience points and leveling mechanics and already runs on over 30,000 sites.

The tools are there. The distinction that matters is between structural gamification, where the course architecture itself supports choice and cumulative progress, and surface-level gamification that just pins badges onto the same old design. Research in Education Sciences in 2024 found that average points-and-badges gamification often just replicates traditional grading’s worst habits. A leaderboard on a quiz is not a sandbox. A course where students level up through mastery, keep what they earn, and choose their path forward? That’s closer to what the games are actually doing.

Checa-Romero and Giménez-Lozano found in Frontiers in Education in 2024 that commercial video games combined with metacognitive strategies enrich academic, social, and metacognitive development. The Crossroads hub in Hades II is a metacognitive space. It forces a pause between attempts. You can’t just mash retry. You have to sit with what happened, adjust your build, talk to people, tend your garden. When was the last time your course design built in a structured moment for students to reflect between high-stakes attempts?

The thing the game knows

Hades II assumes you will die. It builds the entire experience around that assumption. Gee named the principle in 2003. Heeter taught me the frameworks. Kapur published the book in 2024. The research keeps confirming what game designers already understood: failure is curriculum when the environment is designed to make it productive.

The question for educators isn’t whether students will struggle. They will. The question is whether we’ve designed anything that makes the struggle worth returning to, or whether we’ve built a course where one bad grade sends them back to the title screen with nothing to show for it.

One experiment you could try next semester: pick one assessment and redesign it so that failure produces something the student keeps. A revision opportunity. A portfolio entry. A reflection that feeds the next attempt. Give them an Arcana Card. See what happens.

References

Bacalja, A., et al. (2024). Postdigital videogames literacies: Thinking with, through, and beyond James Gee’s learning principles. Postdigital Science and Education, 6(4), 1103-1142. https://doi.org/10.1007/s42438-024-00510-3

Barz, N., Benick, M., Dörrenbächer-Ulrich, L., & Perels, F. (2024). The effect of digital game-based learning interventions on cognitive, metacognitive, and affective-motivational learning outcomes in school: A meta-analysis. Review of Educational Research, 94(2), 193-227. https://doi.org/10.3102/00346543231167795

Checa-Romero, M., & Giménez-Lozano, J. M. (2024). Video games and metacognition in the classroom for the development of 21st century skills: A systematic review. Frontiers in Education, 9, 1485098. https://doi.org/10.3389/feduc.2024.1485098

Faber, T. J. E., et al. (2024). Effects of adaptive scaffolding on performance, cognitive load and engagement in game-based learning: A randomized controlled trial. BMC Medical Education, 24, 935. https://doi.org/10.1186/s12909-024-05698-3

Gee, J. P. (2003). What video games have to teach us about learning and literacy. Palgrave Macmillan.

Kapur, M. (2024). Productive failure: Unlocking deeper learning through the science of failing. Jossey-Bass/Wiley.

What this term taught you about teaching may not surface cleanly in your gradebook or evaluations. If you want a record, now is the time to reflect. After that, the only version you have is what your memory reconstructs in August, and by August, gaps will exist between what you remember and what actually happened.

This post argues for treating end-of-term reflection as instrumentation: the missing measurement that records what evaluations and grades cannot. The protocol is twenty minutes per course, one document, four prompts. The rest of the post is the case for actually running it.

We assign this. We skip it.

You ask your students to reflect. You build it into final projects, capstone deliverables, and the metacognitive prompts you stuck into week six. You grade their reflections, sometimes generously. You tell them that looking back is part of how the learning consolidates.

How often do you do this in relationship to your teaching when it is not required for a report?

The reason most of us skip it is not mysterious. End-of-term is the worst week of the year for reflective work. The practice has no external deadline. And without a structure, “reflect on your teaching” is a prompt vague enough that it does not lead anywhere. So when we try anyway, we default to something that feels like reflection but is not. We default to rumination.

If a student turned in what most of us actually do as end-of-term reflection, you would not pass it.

Reflection as instrumentation

The semester you just taught is data rich. There are hundreds of small observations sitting in your head right now about pacing, about which assignments produced the work you were glad to read, about the moments you saw understanding actually move, about where the AI-use question kept tripping over its own feet, about the questions students asked in the last two weeks that suggested they had been confused since week four. None of that is captured automatically. The course management system does not have it, evaluations will not surface it, and your gradebook only shows the outcome, never the why.

Without structured capture, you have no signal from this term. You only have mood. You cannot make a decent design decision from mood.

The reflection that fixes these runs once, briefly, while ideas are still fresh. The output is not a feeling. The output is a short artifact you can read in July without flinching, and one you will treat as evidence when you decide what to change.

Engineers and scientists trust instruments that are short, defined, and used the same way every time. The reflection that works is the same kind of object.

A protocol short enough to actually run

Run this once per course.

Time: ~20 minutes per course taught Materials: One document Output: Four prompts, answered in plain sentences.

1. What worked, in concrete terms. Name an assignment, a session, or a structural choice that produced student work or student questions you were glad to encounter. Be specific enough that you would believe yourself in three months.

2. What broke, in concrete terms. Name where students struggled in ways you did not predict. Name where AI use changed what your assessments were measuring. Name where pacing missed. The difference between a useful answer here and a useless one is concreteness:

• A useless version reads: Project 2 didn’t land. Students seemed checked out.

• A useful version reads: Project 2, the open-ended AI prompt assignment: most submissions reused the same handful of prompt structures. The constraint was too loose, and the outputs were shallow because students had no reason to push harder than the prompt asked them to.

  The first is mood. The second is data. You will believe the second one in three months. You may believe the first, but it doesn’t offer anything to help you improve.

3. What surprised you. The data point you did not expect. The student question that has been sitting with you. A moment you cannot account for and want to come back to.

4. One question this term left you with. Not a list of ten. One. The question you would still be carrying at the end of August whether you wrote it down or not.

Save the document somewhere you will actually find it. Name it after the course and the term. That is the artifact. You can come back to it in July when you have to make decisions about fall, and you will have something other than your reconstructed memory to make them from.

If you teach more than one course, run the protocol once per course. The protocol is short on purpose. The point is that it gets done.

From signal to decision

The protocol gives you signal. Signal is not the destination. A few weeks after you write the artifact, when the urgency of the term has faded enough that you can read your own answers without flinching, sit with it again and ask the why question. Not what happened. Why it happened. Which design choice produced what worked moment. Which assumption about your students or your assessments was surfaced by broke moment. The why is where next year’s course design actually lives.

Then make one decision. One thing to change for fall, specific enough that you could explain it to a colleague over coffee in September. Not a list of ten. One. Build the artifact for that one decision, finished, before the term starts.

If you do this every term, in five years you have rebuilt your course from the inside, one tested change at a time, instead of running five frantic summer redesigns that mostly cancel each other out.

A side benefit

If you keep these reflection artifacts term over term, you have, by the time any P&T review comes around, a dated record of what you did, what you changed, and why. That record is more useful than evaluations alone for tenure-track and non-tenure-track faculty both, and it is much more useful than what most of us reconstruct from memory at dossier time. We will discuss this more in a future post.

If you want to use AI as a thought partner

A few guardrails first. Use the AI tool your institution has authorized for faculty use. If you do not have one, ask before you reach for a consumer chatbot. Do not paste student names, grades, submitted work, or anything else FERPA-protected into the conversation. The reflection you are running is about your teaching, not about your students. Keep it that way.

Within those limits, AI can do useful work here. It can interview you through the four prompts, asking the follow-up questions that move you from “Project 2 didn’t land” to a sentence specific enough to believe in three months. It can read your draft of the artifact and flag the lines that are still mood, not signal, so you can decide whether to rewrite or cut them. It can sit with you on the why question a few weeks later, after the term has cooled, and push back when your answer feels too tidy.

What it cannot do is run the protocol for you. The point is that you noticed something this term that no one else was in the room to notice. The AI is friction, not author.

One next step

Open a document right now, before the inbox pulls you somewhere else. Name it after the course you just taught and the term you just taught it in. Write three sentences, one under what worked, one under what broke, one under what surprised me. Three sentences total is fine. You can come back to the rest before grades go in.

That is the reflection that will change something next year. The other version, the one you do in the shower, will not.

Further Readings

The post draws on several readings. The works below back the major claims and are the right starting points for anyone who wants to fact-check, dig in, or trace the argument back to its sources.

The rumination versus reflection distinction

Nolen-Hoeksema, S., Wisco, B. E., & Lyubomirsky, S. (2008). Rethinking rumination. Perspectives on Psychological Science, 3(5), 400–424. https://pubmed.ncbi.nlm.nih.gov/26158958/

How memory fails between the event and the recall

Schacter, D. L. (1999). The seven sins of memory: Insights from psychology and cognitive neuroscience. American Psychologist, 54(3), 182–203. https://pubmed.ncbi.nlm.nih.gov/10199218/

Course evaluations: noise, validity, and bias

Boring, A., Ottoboni, K., & Stark, P. B. (2016). Student evaluations of teaching (mostly) do not measure teaching effectiveness. ScienceOpen Research. https://www.scienceopen.com/hosted-document?doi=10.14293/S2199-1006.1.SOR-EDU.AETBZC.v1

MacNell, L., Driscoll, A., & Hunt, A. N. (2015). What’s in a name: Exposing gender bias in student ratings of teaching. Innovative Higher Education, 40, 291–303. https://link.springer.com/article/10.1007/s10755-014-9313-4

Foundational frameworks for reflective teaching practice

Boud, D., Keogh, R., & Walker, D. (Eds.). (1985). Reflection: Turning experience into learning. Kogan Page. https://www.taylorfrancis.com/books/edit/10.4324/9781315059051/reflection-david-boud-rosemary-keogh-david-walker

Brookfield, S. D. (2017). Becoming a critically reflective teacher (2nd ed.). Jossey-Bass. https://iucat.iu.edu/catalog/17047648

Hatton, N., & Smith, D. (1995). Reflection in teacher education: Towards definition and implementation. Teaching and Teacher Education, 11(1), 33–49. https://www.sciencedirect.com/science/article/abs/pii/0742051X9400012U

Korthagen, F. A. J., & Vasalos, A. (2005). Levels in reflection: Core reflection as a means to enhance professional growth. Teachers and Teaching: Theory and Practice, 11(1), 47–71. https://www.tandfonline.com/doi/abs/10.1080/1354060042000337093

Larrivee, B. (2008). Development of a tool to assess teachers’ level of reflective practice. Reflective Practice, 9(3), 341–360. https://www.tandfonline.com/doi/abs/10.1080/14623940802207451

Schön, D. A. (1983). The reflective practitioner: How professionals think in action. Basic Books. https://www.taylorfrancis.com/books/mono/10.4324/9781315237473/reflective-practitioner-donald-sch%C3%B6n

Learning analytics and the limits of LMS data

Selwyn, N. (2019). What’s the problem with learning analytics? Journal of Learning Analytics, 6(3), 11–19. https://learning-analytics.info/index.php/JLA/article/view/6386

Slade, S., & Prinsloo, P. (2013). Learning analytics: Ethical issues and dilemmas. American Behavioral Scientist, 57(10), 1510–1529.https://journals.sagepub.com/doi/10.1177/00027642