There is plenty to be concerned about these days when it comes to AI and privacy. But, when we talk about AI and privacy, we almost always talk in one direction: AI as a force that will supercharge surveillance. A recent story about edtech company Turnitin’s new academic “integrity” product, however, should call our attention to another, equally worrying side to the AI and privacy coin. That is the increasing use of intrusive surveillance technologies to combat the use of AI.

The Chronicle of Higher Education reported this week that with the company’s new product, called Clarity, “Professors will be able to watch a replay of the writing process.” That’s because Clarity monitors typing, revisions, use of copy-and-paste, and time spent writing, all in service of combating student misuse of AI. As one professor quoted in the Chronicle piece noted, it’s the same kind of strategy that malicious actors use to steal sensitive information like bank passwords. As I wrote in a 2020 piece for Forbes, when a new wave of monitoring software was being pushed to combat cheating during pandemic-era remote learning, such tools “border on what might be called ‘spyware’ in the cybersecurity world.”

Turnitin isn’t against students using AI, however. They just want students locked in to using their AI. Clarity also bundles an “AI assistant” that gives students feedback and an AI citation checker. That’s right, one of the anti-AI spyware’s selling points is its own inclusion of AI.

This might all seem like a contradiction, but it’s not. As someone who has studied the rhetorical tactics and logics of national security discourses, I find the patterns here familiar. They include the use of fear appeals, moral panic, and the logics of emergency and exception to justify the expansion of surveillance in the name of safety and security. And the people causing the danger, stoking the fear, and selling the solution? It’s often the same people. That’s called a protection racket.

Unfortunately, when it comes to fighting AI with surveillance while also selling AI, Turnitin is not alone. To be in compliance with EU AI Act watermarking requirements, some big players like Anthropic got in on the act this week too. This emerging AI-powered AI protection racket is what we’re focusing on in this piece.

The Overlooked Direction In The AI-Privacy Relationship

Let me start by acknowledging the very real threats to privacy from AI’s use as an engine of surveillance. This includes the proliferation of automated license plate readers (ALPR), surveillance cameras, and the machine learning algorithms turned loose on all of our personal digital data for years now. This is to say nothing of what the cloud model providers like OpenAI and Anthropic are learning about us all, how the data is being protected (or not), and how it might be misused. (And lest you think I’m downplaying the threat because I just want to promote AI slop, you can go read my academic research on the matter first and then get back to me.)

My concern in this piece runs in the reverse direction and focuses on surveillance and the erosion of privacy undertaken in order to combat AI, and legitimated by anti-AI sentiment rather than by pro-AI enthusiasm. These moves are not framed as surveillance, of course. They are about “integrity,” “protection,” “transparency,” and “safety.” But at bottom, they are surveillance like any other and the “transparency” is for thee, not for me (i.e. the providers of the tools).

There is no shortage of such tools. We started with the most obvious, AI detectors. If you’re on Substack, you are no doubt aware of Pangram. But it’s only one among many, like Copyleaks, GPTZero, ZeroGPT (the naming is very original), Originality.ai (irony alert), Scribbr, and many, many more. Some, but not all, of these also offer AI “humanizer” tools meant to help you defeat their own detectors.

Repeat after me: Protection racket.

Others, like Turnitin, go a step further. In addition to AI detection, they offer other forms of “integrity”-enhancing technologies that function like certain kinds of malware (e.g. RATs and keyloggers). Turnitin’s Clarity boasts keystroke telemetry, revision logging, and screen replay, for example.

Once again, however, Turnitin is not alone. During the pandemic, we saw an explosion of schools requiring students to install intrusive exam proctoring software with access to the student’s webcam, microphone, screen, and browser history. In some cases, they also required the student to submit a scan of their room and biometric identifiers, and to submit to eye tracking via their webcam.

Biometrics are once again on the menu of surveillance tools for those wishing to combat AI. In April, the BBC reported that “Users of the dating app [Tinder], as well as other major platforms such as video calling service Zoom, will be able to scan their irises to earn a ‘proof of humanity’ badge attached to their profile or name.” To earn your humanity, you only need to “submit to a scan” by an “orb-shaped scanning device run by World network.” What is “World network”? Oh nothing, just

“part of Tools for Humanity, a start-up co-founded and chaired by Sam Altman, who is also the head of ChatGPT-maker OpenAI.”

“Once a person is confirmed as human by the technology they receive a unique identification code, which is stored on their smartphone and considered their World ID.”

Why would anyone be against “tools for humanity”? AI slop is a threat and something must be done! I’m sorry, but if you can’t see that we need Sam Altman to issue us all a mark of the...uhh, I mean, badge of humanity so that we can buy and sell and converse safely and transparently, it’s probably because you have something to hide! Sign up today. World ID #666 gets a free year of ChatGPT Pro.

The hypocrisy and stupidity of it all would be hilarious if the stakes were not so dire. Mass surveillance, biometric capture, erosion of privacy, and automated judgment of individuals are among the AI risks that command the most public concern. And rightly so! And yet...every single countermeasure in the inventory above is mass surveillance, biometric capture, erosion of privacy, or automated judgment of individuals. The defense and the threat are the same thing pointed in different directions, manufactured and sold by the same damn people!

Repeat after me: Protection racket!

The Rhetorical Invention Of The AI Protection Racket

In classical rhetoric, “invention” is the art concerned with finding and framing the available arguments for a case. Cicero defined it as “the discovery of valid or seemingly valid arguments to render one’s cause plausible.” Another translation of Cicero says, “Invention, is the conceiving of topics either true or probable, which may make one’s cause appear probable.” Invention proceeds by topoi, which are the commonplaces, the stock of argument forms a speaker searches rather than originates. (For all of you looking to the humanities and rhetoric and the like for the authentic human voice and for a defense of originality, I’ve got bad news...but I digress.)

Though in the classical tradition inventio was connected to the individual speaker or “rhetor,” we see a hint in the second translation of the more modern understanding of the term, which sees invention as an outcome of collective action, the result of networks of discourse interacting, morphing and mirroring one another over time. In our book, The Rhetorical Invention of America’s National Security State, this is the sense of “invention” that my coauthors and I mean. We write,

“the production of a national security state is a communicative achievement, where technical, political, and cultural arguments are marshaled together to get the warranted assent of U.S. audiences who come to believe they need to live in securitized states” (p. 2, emphasis added).

This is not the work of one lone actor, however. The national security state has “many authors and countless purveyors”—including its own subjects—and is not something merely “conjured by lone ‘imperial presidents’ abusing ‘unitary executive’ theories of state power” (p. 2). (That is, such actors play a role, but they are not the only ones.)

So what does this have to do with AI and surveillance? Two things. First, because invention is a team sport, rhetoric is subject to “drift.” Topics and argument forms deployed in one context drift to other contexts over time. Second, “safety, “security,” “protection,” and their commonplace arguments about “emergency” and the “necessity” for “exceptional” response measures are particularly viral in their ability to drift from one context to another.

In discourses of security, emergency, and exception, the protection racket plays a starring role. I wish I could take credit for the idea, but it does not belong to my coauthors and me. Sociologist Charles Tilly used it as early as 1985 and critical security studies scholar Mark Neocleous has further developed it. But in our book, my coauthors and I took particular note of the feedback loop involved in the process, and it’s important for the discussion of AI-powered AI protection.

“just as the national security state self-authorizes its members, demarcating the space of national security expertise from non-expertise, so too does it produce and reproduce the ‘needs’ and ‘desires’ for the ‘solutions’ that only its members can provide. That is, the national security state not only produces security, but also insecurity. In an ironic twist, both the national security state’s successes and failures can result in a positive feedback loop of seemingly ‘needing’ and desiring yet more security.“ (p. 15, emphasis added).

Sound familiar? Turnitin’s detector fails, and the remedy sold for that failure is more Turnitin. AI allows realtime deepfake video; Sam Altman has a proof of humanity badge to sell us for the small price of storing our iris in his database. Failure and success both generate demand. Sure, there are plenty who are selling surveillance-as-security. But there are just as many demanding it.

Destroying The Village To Save It

My coauthors and I were writing in the midst of the U.S. “war on terror” and the Snowden revelations about NSA mass surveillance. We were paying particular attention to how elites sold “security” and how citizens demanded it using rhetorics of emergency and exception, as well as how those of us who questioned the wisdom of this crowd were often vilified in the process. You’re with us or you’re with the terrorists! If you’ve done nothing wrong, you have nothing to fear! What do you have to hide?

We can see the same rhetorical moves in the AI debate today.

• AI is an existential threat! AI will destroy education, writing, creativity, etc.! (threat inflation)

• Something must be done now! (emergency/necessity)

• False positives and loss of privacy are the price we must pay! (necessity/exception)

• You’re either for or against AI! “Nuance” is naivety or excuse! (black and white thinking)

• If you’re questioning, it’s because you’re cheating and trying to hide it! (vilification/othering)

The vilification and othering found in the anti-AI discourse, including ableist and dehumanizing rhetoric directed at some AI users, is troubling enough. But like all such fear-based rhetorics, it is also leading to a self-fulfilling prophecy.

Again, from my own work on national security discourses, I’m reminded of the case of airpower. In brief: Pre-WWII theories predicted that aerial bombardment against an enemy’s key industrial nodes would lead to rapid victory despite plenty of early evidence to the contrary, which was largely explained away or ignored. When the theory met reality and was proved false, bombardment was directed even more heavily at civilians. The strategists doubled down. In an effort to foil an enemy’s aerial attack, civil authorities instituted precisely the electrical blackouts that airpower theory predicted would lead civilians to panic and demand their political leaders sue for peace. Luckily for the Allies, that didn’t happen either. But in the process, efforts to protect arguably led to the very harm that leaders were trying to prevent.

As I noted in my book on the use of doom rhetoric in cybersecurity discourse,

“One study of the mental health effects among victims of strategic bombing found that excessive precautionary measures taken in an attempt to prevent the panic and paralysis predicted by theorists did more to ‘weaken society’s natural bonds and, in turn, create anxious and avoidant [sic] behavior’ than did the actual bombing.

[...]

“fear of what might happen to society were the power grid to fail led to a self-inflicted lack of power that not only did not have the desired military effect but may also have been an example of excessive, counter-productive precaution that contributed to widespread anxiety” (p. 103).

Military historians, disaster sociologists, and historians of technology have explored the effects and effectiveness (or lack thereof) of strategic bombing in great detail.

The analogy might seem a stretch, but it’s not. Again, the general argument forms drift from context to context.

1. A predicted collapse. In the case of airpower, of society and its bonds of trust. In the case of AI use, of academic integrity, of authorship, of the capacity for critical thinking, of the ability to know what’s real or not, etc.

2. Contrary evidence explained away or ignored. Early cases of failed bombing that should have caused a more nuanced assessment were ignored or explained away. Similarly, those who provide detailed explanations of their own, nuanced and careful uses of AI are ignored, accused of lying, or accused cheating. Studies complicating the picture about AI and critical thinking and cognitive offloading are published, but the same tired studies confirming the anti-AI stance continue to be cited.

3. Precautionary measures adopted at scale. Technological solutions, from self-imposed blackouts to self-imposed digital surveillance, based on little to no evidence of effectiveness.

4. Self-imposed damage. The precautions producing the very damage they were meant to prevent. Blackouts that caused anxiety while failing to prevent enemy bombing. The erosion of trust between teacher and student, degraded writing environments in the name of enforcing “integrity,” false accusations of cheating with real academic consequences, and of course, relying on AI-powered surveillance to protect against the supposed dangers of AI.

Conclusion

The impetus for this piece when I started it last week—once again, I’m failing miserably at using AI to do my thinking and writing for me—was the news of Turnitin’s academic “integrity” malware. In the meantime, Anthropic announced its EU compliance by way of watermarking. The next “drift” for AI “protection” is into law and technological enforcement.

Pangram and Turnitin, you can decline. A legal mandate with global force? Not so much. This is all the more true when the enforcement becomes technological infrastructure built into the system itself. Pangram relies on someone deciding to scan a piece of writing. Turnitin’s writing replay happens on a teacher deciding to watch a student work retroactively. But a Claude watermark creates a document that tells on itself and you, permanently, to everyone who encounters it, whether you did anything wrong or not. No human accuser or judgment call is required. Where the combating of AI is concerned, the human has been removed from the loop, with claims and judgments offloaded to the machine.

The collective nature of rhetorical invention reminds us, however, that this was not done to us. Many faculty asked for it. Many administrators bought it. Democratically elected leaders in the EU mandated it. Nobody was duped. People wanted a machine to tell them who “cheated.” But every countermeasure on the menu is the thing we said we wanted to avoid.

That same fact can be read the other way and can provide some solace, however. Rhetorical invention runs on assent, and assent can be withdrawn. None of this is inevitable, including the legal mandate. Statutes are written by people and can be written differently.

The first step is recognizing what is happening. Pangram, Clarity, and Claude watermarks are all instances of the materialization of a set of collective beliefs about AI and its possible futures. But this does not happen on its own. Richard Doyle has included “rhetorical software” among those technologies that help to enact the “materialization of belief.” Writing about the history of the life sciences, he says, “Rather than a mere description or heuristic for the life sciences, the rhetorics of code, instruction, and program materialized beliefs into sciences and technologies” (p. 3).

Like software, such rhetorics are most powerful when they go unseen, unthought, when they disappear into the background. As Lloyd Bitzer noted in his account of the “rhetorical situation,” the rhetor “not only harnesses constraints given by situation but provides additional important constraints,” such that discourse not only shapes what can be seen, said, and done, but also the very subjectivities of “speakers” and “audiences”—us, them, good, bad, pro-, anti-, honest, cheaters. As Doyle argued, even those caught unawares will be affected as “rhetorics . . . pass through fields and agents as intertextual forces that recast knowledges and their knowers while sometimes remaining in the realm of the unthought” (p. 3).

But when rhetorics bubble to the surface and become a topic of their own, that is when change is possible. Surfacing the problematic rhetorics and assumptions powering the emergence of AI-powered surveillance as an AI protection tool is the first step to averting a self-fulfilling prophecy.

So, one last time, repeat after me:

Protection racket.

But wait…there’s more!

Pangram inside of Substack has granted my a proof of humanity badge. Great tool they have their. It’d be a shame if someone broke it.

Let’s start with what this post could have been. It could have been a confirmation of something I already believe: AI detectors are unreliable and people will be harmed when unfairly and falsely accused.

I still believe that. But, the main focus of this post will be about a funny thing that happened along the way to not writing that other post — a “worked example” of applying those “critical thinking” and “humanities” skills so many talk about and advocate without demonstrating.

When AI Tells You What You Want To Hear

This morning, I watched a livestream about creating a “Content Radar” system to find items of potential interest to write about, have AI score them, and make suggestions. It’s almost impossible to keep up with all the topics I cover in my university teaching, research, and Substack writing. So, I figured, I’d follow along and give it a shot.

I created the system and ran it and immediately got the following results.

Turnitin admitting its false positives are higher than previously admitted? 880,000 papers unfairly flagged? That does sound like something to write about given the recent Pangram controversy.

But it’s times like this that we need to be extra careful. It’s especially tempting to not check the sources as carefully when they agree with you.

Close Reading & Evidentiary Reasoning

I should back up and say that the screenshot above is actually from the second run of the Content Radar. The first run included the Turnitin statistic in a hallucinated headline. I clicked the headline, but the link took me to “Schools Are Catching the Wrong Kind of Cheating.“ That headline doesn’t even mention Turnitin. I clarified that the agent should only apply links to actual headlines, not inferred facts and certainly not make up headlines. What you see above is the result of the second run.

Ok, but still, what’s it got to do with Turnitin? The opening of the piece says,

In June, I attended a three day conference outside Orlando where over a thousand educators talked about artificial intelligence. The keynote speaker, C. Edward Watson, projected some interesting statistics generated by Turnitin, the plagiarism service many schools use regularly. In its first year of AI detection, Turnitin reviewed more than two hundred million student papers, flagging twenty two million of them as significantly AI-generated. The company had also quietly revised its false-positive rate from one percent to roughly four. Watson did the multiplication out loud: 4 percent of 22,000,000 is 880,000. In a single year, the tool we bought to protect academic integrity told schools that 880,000 pieces of honest student work were deemed fraudulent.

There it is. That’s the statistic that the AI “content radar” system had pulled out. Now we’re in business!

Well, not quite. Here are my next questions:

• Is there a link to a source supporting this claim?

• Who is the author and what is this publication?

• If the author’s attendance at a speech by C. Edward Watson is the only source, who is this Watson fellow?

• This seems to imply that these statistics are recently revealed, but that is not stated explicitly. When were these figures released?

I know. I know. When you write it all down like that, it seems kind of paranoid to have so many questions after reading one paragraph. But these really are the kinds of questions I automatically ask, even if I’m not doing so explicitly and in a bulleted list. It’s just a habit at this point. Why?

As I have mentioned before, I learned much of this when I was a competitive debater during high school and then coach (briefly) as an undergraduate. You learn early that you better have your ducks in a row before you walk in and start reading evidence from shoddy sources. Your opponent will find every possible weakness. Of course, since then I’ve honed those skills through the remainder of my education and career as an academic researcher and educator, all of it in the humanities.

Let’s go through each of the questions above.

Share

Is there a link to a source supporting this claim?

There is no link in the paragraph mentioning the Turnitin statistic of interest. I scan the remainder of the article and there are no links at all. Not a good sign. It’s 2026. Linking to sources is easy and should be expected.

Ok, so, next question. If the author and publication are of sufficient knowledge and authority, maybe I can cite them without finding their original source.

Who is the author and what is this publication?

The author is Jason Robinovitz, not a name with which I’m familiar. He is listed as “Co-Founder at National Test Preparation Association” and “Chief Operating Officer of Score At The Top Learning Centers & Schools, a family owned group of tutoring companies and schools in South Florida.” Again, the first is an industry group with which I’m unfamiliar. The second is his own family business. Neither is much for me to go on without further digging.

Solutions Review seems to be a generic IT industry blog and news aggregator. The original publisher, Insight Jam seems to be something similar but the link goes to a signup page, not the original article. I’m not signing up. Dead end.

To be clear, none of this means Mr. Robinovitz or these sources aren’t credible. But it does mean that I am not certain enough of their credibility to cite them as trusted authorities if I want to repeat the statistics about Turnitin. These are not news sites and Mr. Robinovitz is not a reporter. He did not independently uncover and verify that statistic. By his own admission, he got it from C. Edward Watson.

By the standards of academic citation of evidence, this article is not a citable source for the claimed statistics. At best, I could say something like “some have claimed that Turnitin revised its statistics” based on this source. But I cannot say confidently, only based on this source, that it definitely happened. I also cannot say when it happened.

Robinovitz says he got the statistic from Mr. Watson. So...

Who is this Watson fellow?

So, off to SearXNG I go. (You are running SearXNG locally, right? Right?)

Turns out Dr. Watson is “Vice President for Digital Innovation with the American Association of Colleges and Universities (AAC&U).” Now that’s an organization I recognize. This tells me that Dr. Watson has some expertise in this area and is in a position to know such details about Turnitin. My confidence in the claimed statistic has now gone up.

But, it’s still not citable. The original source is still not strong enough on its own. And citing it because Robinovitz said that Dr. Watson presented it is a combination of second-hand evidence and appeal to authority. I didn’t hear Watson make the claim. There’s no link to a video or transcript of him making the claim. And even if he did, if he did not provide a source, then I’m citing based on his word. He seems like a stand-up guy, but he doesn’t work for Turnitin and without some additional source, I cannot confidently cite him either.

Back to the drawing board. Next question.

Can I find independent verification of these statistics and when they were released?

This next step is where I turned to an AI tool that I created, have used for over a year, and that I trust. It is my “Claims Evidence Finder” project in Perplexity. Here’s what it does.

I enter a claim, hit enter, and it finds three debate card-style pieces of evidence in support of the claim. But only if evidence exists. If it cannot find credible evidence, it tells me so. To make this happen, it combines a detailed system prompt with a small knowledge base of reference documents to help steer the model’s output.

I use this tool regularly in two scenarios. First is one in which I am fairly certain that a claim is correct or that someone has made a claim (regardless of whether it is correct). The sky is blue. I need something credible to cite. Or, I know I’ve seen people claim that the sky is blue and need a few examples of such to point to.

The second scenario is the one that played out here. This is the one where I come across a claim made without proper citation of supporting evidence. The Claims Evidence Finder serves as a sort of verification/fact checker tool. I copy in the text of the claim and hit enter.

Two things happened in this case. One is seeming confirmation of the statistics cited by Robinovitz. However, the big red flag here for my post-that-could-have-been was seeing that these numbers were from 2023. These might be real, but this is old news.

Here is what an example “card” looks like. It includes the “tagline” (paraphrase), full citation, link, and quote supporting the claim.

Of course, the next step is to verify the verification. Rule # 1: Treat AI like an arms treaty with the Soviets. Trust but verify. I clicked the link and put Ctrl+f to work again.

The included quotation is indeed found in its entirety at the provided link and includes the key statistic in question.

If I wanted to use the other two cards, I would check them too. But at this point, I had what I needed. I was convinced that the stat is real. Inside Higher Ed is a trusted trade publication for higher education. What’s more, they link to the original blog post on the Turnitin website. But I was also convinced that this wasn’t new information that warranted writing up a post. Ah, the one that got away.

Share

Critical Thinking And/With AI

Let’s return to trust but verify. Because I’m a former debater and an academic, I can always already hear my opponent and/or Reviewer 2 coming up with counter arguments. So I can hear some of you saying, “But you used AI to fact check! You trusted it! But but but...it hallucinates!”

I used a tool that I created and which instantiates a set of practices and formatting that I have used since 1992. As I wrote in a previous post, it’s a method that goes back to the Ancient Greeks. I did not trust AI as much as I trusted myself, my way of doing things, and the Western humanistic tradition of the florilegium, commonplace book, and zettelkasten. That is what the AI tool I created is doing. And I know whether it’s doing it correctly or not because I’d been doing it myself for over 30 years, first on notecards and loose sheets of paper, then in various electronic database configurations up until the present.

I also know whether it is doing it correctly or not and to what degree I can trust it because I’ve been using it and refining it over more than a year. I originally set up the same system in Perplexity, Gemini, and ChatGPT. At first, ChatGPT was most reliable. Then GPT-5 came along. It broke and hasn’t worked since. Gemini never worked. It was always a hallucinating mess and never followed the formatting rules consistently. Perplexity started out middle of the road but is now the sole home for the Claims Evidence Finder. Refining the system prompt over time, combined with improvements to the models I use with it (usually Claude models, but recently GLM5.2 on occasion), have resulted in the Perplexity version being consistent and trustworthy.

Nonetheless, because it still does occasionally hallucinate a portion of a quotation or mashup noncontiguous text into one quote, I always visit links and check exact wording in context before quoting and citing what it provides. The number of incorrect quotes has gone down over time though and it’s rare to find one these days. But, again, you have to verify.

Conclusion

So, that’s it? I wrote a whole post about checking one claim from a new AI system...and in excruciating detail? Yes. Because we keep hearing about the need for critical thinking, humanities skills, and judgment. But few show us what that looks like in practice.

This is what it looks like. None of it really required the use of AI. As I said in my post on my note-taking process, I use AI sometimes to augment or amplify what I already do. Here, I used good old fashioned critical thinking and verification skills to evaluate a new AI system’s outputs, then to verify the facts in the source the AI cited. Doing so revealed that the original AI output and the source on which it relied were not necessarily wrong, but rather, misleading to a degree that I abandoned my original post idea.

I used AI to assist in the verification process, but following methods with which I am well familiar and using a tool I have developed, evaluated, and refined over more than a year’s worth of use. I did not give a chatbot the vague task to “fact check” and then just trust what it said. I trusted my own process, as instantiated in an AI tool that I built and evaluated myself. In the end, however, the AI tool was not necessary, merely a nice convenience.

AI-in-the-loop, human in charge. This is the way.

If you would like to learn a straightforward method of fact checking and verification very similar to what I described above, take a look at the SIFT Method. It’s the first link in a list of resources that librarian Hana Lee Goldin posted, as if on cue, while I was writing this post! See the full list HERE. Also, be sure to subscribe to Hana’s Substack.

Editing the post in Obsidian with NotebookLM opened in an internal Obsidian browser tab.

If you’ve been paying even a little bit of attention on Substack the last few days, you will have noticed the absolute sh¡tstorm that followed the platform’s announcement of a deal with AI detection company Pangram to allow readers to scan posts, notes, and comments for AI use. In this post, I’m going to show you the behind the scenes of how I used AI to draft my last post and how I got around Pangram.

But before I do that, some background...

Playing Cop with a Broken Radar Gun

I’ve made no secret of my thoughts on the matter. To say I think the people cheering this are silly would be an understatement. This is a platform company with, let’s be honest, some ethical problems of its own, making a deal with an AI company to give those concerned with “AI slop” an AI tool that tells them what not to read. It’s the very outsourcing of judgment and critical thinking that I think we can all agree must be avoided.

I remain a “skeptical enthusiast“ when it comes to AI, torn between curiosity and fascination on one hand and deep concern on the other. But that means I don’t avoid the use of AI. Instead, I am using it daily, in lots of different forms. I’m currently working nonstop to launch a 16 week course titled, “AI Literacy for the Modern World,” for upper division undergraduates at the university where I’ve taught for 18 years. To do that effectively, credibly, and ethically, I have had to use this technology to learn about it.

In the process, I’ve done a lot of experimentation with incorporating AI into my normal work, from writing to course preparation. Some experiments have gone well. Some have been so-so; others failed. But at the end of the day, I will be the one to decide, for myself, based on my own decades of experience “doing the thing”--publishing books, peer reviewed research, and online commentary--whether my use is valuable and ethical. (You like that em dash? I decided that. But even if I didn’t, I decided to keep it.)

Some of the experimenting I’ve been doing is with AI detectors because, yeah, if you haven’t heard, students have been using AI to cheat, detectors are one response, and they are themselves another use of AI that is controversial. So, a year ago I got a subscription to Copyleaks. A few months ago, I signed up for Pangram.

I’ve used them to scan student papers. But I have not failed or even deducted points from a student as a result. Why? Because I know that they are unreliable both from reading about their unreliability and from testing them out myself (we’re getting to that part, I promise).

Instead, I showed the student the results. I suggested maybe they had used Grammarly or some other tool. I warned them that such tools can result in a positive AI detection score and told them to be careful. I showed them screenshots of my own comments to them getting flagged as AI because I had dictated them using Apple dictation and cleaned them up using Apple’s built-in writing tools. Screenshots included the original, raw dictation, the cleaned up version, and the AI detection score on the latter.

Had some of them actually cheated? Almost certainly. But I couldn’t prove it and as I’ve gotten older, I like to avoid “being the cops” in my relationship with my students. So, I gave some grace (maybe undeserved) and made it a “teaching moment.” And I slept just fine.

Share

An AI-Powered Substack Post Speed-Run

I know the detectors are flawed because, well, I’ve played with them. When I say “play,” I mean like a cat slowly drowning a mouse in its water bowl. What if I change this phrasing? Bonk, splash. How about shorter sentences. Bloop, bloop. Short then long? Gurgle. Where does this thing break down? How can I cause it to fail? How can I cause it to tell me AI text is human and human text is AI? We’ve got a floater!

So, here’s one example of that use of the supposedly super accurate Pangram, which has even been used to accuse the Pope of posting AI slop. (You’d think that would’ve been the definitive, shark-jumping moment. But no.)

As mentioned above, I used AI to help me write my last Substack post about the hype and doom nonsense around Mythos. Here’s how I did that and how it ended up with a 100% human score. I used a combination of NotebookLM and local AI, combined with good old fashioned editing and -gasp- human-fingers-on-keyboard writing too.

Step 1 - Context

One of my favorite AI tools, NotebookLM, was the core of the operation to speed run the AI detector and infiltrate your minds with my slop. Muwahahahaha!!

I am an inveterate information hoarder and notetaker. And I’ve been following cybersecurity related hype and doom narratives for over a decade. So, you can bet your Fable tokens I had collected and stashed away a bunch of articles and written a bunch of thoughts on the Mythos kerfuffle.

So, step one was to create a new notebook and upload all that material as the source base from which I would let AI do all my thinking for me.

Context given to NotebookLM included relevant news articles, statements from Anthropic, a YouTube video from Cal Newport’s podcast, my book on cyber-doom rhetoric, one of my peer reviewed articles on the subject of cyber Pearl Harbor rhetoric, and a personal research memo on the confluence of AI hype and cyber-doom rhetoric from my Obsidian vault.

Or...maybe not.

Step 2 - Tell AI What To Say

You thought you had me. This fool let the AI write it all! I knew it!

But nope. That’s a negatory, pig pen. It’s not X, it’s Y.

I knew what I wanted to say because 1) I’d been reading and writing on this particular topic already and 2) I’d already written an academic book on the broader topic of cyber doom rhetoric. So, you know, who better to tell me whether the AI’s output was good than...me?

So, I gave NotebookLM a prompt that told it what we are drafting (a newsletter post), what my argument was, and the key points I want to include to support that argument (drawing, of course, on the sources I’d added, which included things I had actually read/listened to and my own notes.)

And with that, the robot dutifully went off and wrote the post you all read--you did read it, right??--and I collected my big Substack check.

Step 3 - Tell AI How To Say It

Not quite.

It did make the assigned argument, with the assigned supporting points, and citing the provided sources. But it sounded like robot vomit. It was not usable in its current form. So, I wondered, could I get it to generate something that sounded closer to my writing style? And by doing so, could I fool the AI detector?

I put on my mad scientist hat on and went to work.

Step 3a - Teach The AI How To Write

I turned on the lights in my home AI laboratory, a MacBook Pro with local LLMs installed, and fired up Open WebUI. I proceeded to create a knowledge base (i.e. a local vector store readable in Open WebUI) with almost all of my books, journal articles, op-eds, etc.

I fired up my largest local model, gpt-oss:120b, pointed it at the knowledge base, and asked it to analyze the writing style and write up a system prompt to instruct a model to write in that style.

First pass at having a local model create a “write like me” system prompt.

I had it generate an example output using the prompt. It was OK, but it was a little too bland. My writing has become a little more spicy (a.k.a. polemical) over the years and I wanted that reflected in the prompt. So, I asked for a revision.

Local model response to a request for a more polemical but still academic and rigorous style.

I asked the model for an output using the writing style system prompt it had created and still drawing from the knowledge base of my own writing.

Prompt to gpt-oss:120b to use the system prompt it created and the attached collection of my writing to produce an output that sounded like me.

The Prompt:

Demonstrate the updated prompt by writing an introductory paragraph for an essay that uses the rhetorical tacts of cyber-doom discourse to critique the use of the same tactics in narratives of AI doom.

The Response:

The discourse of “cyber‑doom” has long relied on a bundle of rhetorical tactics—hypothetical catastrophes, war‑and‑disaster metaphors, and an urgent “act‑now” imperative—that transform ordinary technical vulnerabilities into existential threats and thereby justify the expansion of state power or vigilantist interventions [1]. By foregrounding imagined futures of societal collapse, these fear appeals sideline empirical evidence and instead mobilize panic, paralysis, and calls for pre‑emptive authority [1]. Contemporary narratives of “AI doom” recycle the very same playbook: they invoke speculative apocalypses, amplify the specter of loss of control, and marshal the language of war to frame artificial intelligence as a looming, ungovernable monster [1]. This essay argues that the uncritical transplantation of cyber‑doom tactics into AI discourse not only distorts the actual risk profile of emerging technologies but also reproduces the authoritarian impulses and policy overreactions that the original cyber‑doom rhetoric has historically enabled.

I was surprised, to say the least. To me, it sounded like me. It accurately represented my writing on cyber-doom rhetoric, its perils, and applied it to AI doom in a way with which I agreed.

So, I decided to do the next natural thing. I did some non-IRB approved human subjects research. I sent the output to a co-author and good friend (who happens to hate AI, btw) and said I was working on an essay and asked how this draft introduction sounded. He said it sounded good. I told him it was AI that I had told to write like me based on my work. He was not thrilled, of course, but did admit it sounded like something I would write.

Step 3b - Defeat The Detector, Profit!

It sounds like me. And last time I checked, I’m human. Off to Pangram I went. I pasted in the text and... -drumroll- 100% AI.

Back to the lab. I prompted gpt-oss again, “Good try. This text gets flagged as AI written. Can you add some more burstiness, perplexity, and generally make it more humanized?” I’d heard those terms used by YouTube AI “humanizer” gurus. Maybe they would unlock some secret sauce inside the model. If the secret sauce is writing that sounds awful, then yeah, the magic words did the trick.

Back to Pangram. Same result. But now with text I didn’t like and didn’t sound like me either. I’d hit the limits for this step, I decided.

Subscribe now

Step 4 - All Together Now

I took my writing style system prompt back over to NotebookLM, pasted it into the overall system prompt setting, and ran the original prompt again that told it what I wanted to argue, supporting points, etc. The result was better, but not by much.

System prompt added to the chat configuration settings that instructs the model to write in my style.

You are an academic writer who produces scholarly prose in the style of the source named `Sean Lawson - Cybersecurity discourse in the United States - Cyber-doom rhetoric and beyond.pdf`. You may adopt a polemic, critically challenging stance when appropriate.

• Write in a formal, evidence‑based tone; avoid colloquialisms, but you may use assertive, evaluative language (e.g., “contradicts,” “undermines,” “exposes,” “fails to account for”).

• Structure your argument logically: introduce the issue, present the dominant discourse, then critically interrogate it by exposing contradictions, hidden assumptions, or power relations.

• When you challenge a position, do so with clear justification: cite data, prior scholarship, or logical analysis. Every critical claim must be explicitly referenced in a consistent citation format (APA, Chicago, etc.).

• Feel free to employ rhetorical devices such as strategic questions, contrastive framing, or brief counter‑examples, provided they serve the scholarly argument and are grounded in evidence.

• Focus on rhetorical and material analyses of security, surveillance, and technology and the attendant ethical concerns (see, e.g., the materialist reading of Snowden‑era surveillance ; second‑order methodological critiques that reveal power relations ; and systemic media‑literacy challenges).

It occurred to me that the gpt-oss output was probably better not just for the prompt but also because it had all my writing it could look at along with the prompt. So, I added a few of the most relevant items to the Mythos topic to the NotebookLM source list and tried again.

The final chat prompt to produce the post draft, including explicit instruction to use my writing to guide the analysis and writing style.

Draft a newsletter post that argues that the U.S. government's action requiring Anthropic to remove access to its Fable 5 and Mythos models demonstrates the dangers of the increasing convergence of AI doom rhetoric with cyber doom rhetoric.

- Explain what happened with the government requiring removal
- Explain how this follows hype from Anthropic about the supposed dangers of its Mythos model and its early decision to withhold public release
- Explain how Anthropic's hype is exemplary of a recent convergence of cyber-doom and AI doom rhetoric
	- Define cyber-doom and its rhetorical characteristics
	- Provide examples of how AI and cyber-doom rhetorics have merged recently

- Explain the dangers of doom rhetoric based on the Lawson source about cyber-doom and the Lawson and Middleton source about cyber Pearl Harbor. In particular, explain how military/war/security framing can encourage government overreach and overreaction, which is what we see in the case Mythos.
- Conclude with recommendations for how to talk more productively about the cyber risks posed by AI.

Throughout, you should base your analysis and your writing style on the sources: `Sean Lawson - Cybersecurity discourse in the United States - Cyber-doom rhetoric and beyond.pdf` and `Lawson and Middleton - 2019 - Cyber Pearl Harbor Analogy, fear, and the framing of cyber security threats in the United States, 1.pdf`. These sources are written by you, so they form the foundation of your thinking on this topic and represent your style and tone when writing.

Share

Step 5 - Fool The Robot Cops & Their Minions...For Real

The results were much better. The output sounded more like what I’d hoped for. It hit all the main points. Surely, this would bypass the detector! I highlighted the text in NotebookLM, right clicked, ran Pangram, and...

Pangram AI detection before editing.

Womp womp!

Still 100% AI. Makes sense, I suppose, because it was 100% AI! (Well, maybe. More on that in a minute.) But it also wasn’t yet what I wanted it to be either. I therefore proceeded to run a very sophisticated black hat humanizing process on the AI slop.

Editing. I edited it. I moved stuff around. I deleted some. I added a lot. I changed even more. I did the thing you do with a rough draft. Any rough draft. NotebookLM’s’ original draft was 659 words. The final, published version: 1,744 words. 1,085 original human words added to the 659 carefully prompted and steered AI words.

And the final verdict:

Pangram AI detection after sophisticated black hat humanizing--a.k.a. editing.

🎈💰 Ed McMahon and the Prize Patrol showed up with balloons and a giant check!! 🎈💰

Subscribe now

What’s It All Mean?

Hell if I know. And, like, that’s kind of the point. You don’t know either. The people howling the loudest? Definitely not.

I’d venture, though, at some level, both scores are a lie. Was the original, AI text 100% AI generated? Kind of. It’s closer than the final text. But did the AI prompt itself? Did the AI tell itself what to write, in what order, in what style and tone, and which sources to cite in the process? No. What percentage of the production do we place on the research, note taking, ideation, and direction? I don’t know. But I do know that the AI did not wake up one day in May and decide to do all that on its own and then just go do it.

Similarly, however, is the final version 100% human as scored by Substack’s favoritest new robot inquisitor? Definitely not. I just showed you it isn’t. Some of those words that are still in the post were definitely written by AI.

But are they my words? Yes. Why? One, because I put in the work and care to “produce” them, through a combination of methods, typing or otherwise. I labored over them. I used care in engineering the context, prompting, editing, and good old fashioned writing too. I read every one of those words multiple times, over many hours spread over two days before I hit publish. I stand by them. If they are wrong, that’s on me.

In short, they’re mine because I said so. Because I authorize them--even if I don’t author them--in my own name.

Conclusion

If you made it this far, it’s probably because you weren’t actually in it for what the clickbait title seemed to promise. And if you were, sorry. You didn’t get the magic bullet to bypassing Pangram. The answer in this n=1 experiment ended up being good old fashioned research, reading, writing, and editing. Sorry. I’m still an academic and that’s what I know.

The bigger objective here, however, was to complicate and simplify the whole thing simultaneously. Neither the use of AI, nor AI that detects AI, is as straightforward as many would have you believe. Effective use of AI is not a matter of type a prompt and get an answer. Better or even best AI outputs are still not enough; but they’re not nothing. Where the human’s role ends and the machine’s begins in creative labor has never been an easy question--and it has always been a question, with every new media technology. Generative AI is but our latest spin on the technopanic merry-go-round.

But therein lies some solace. We’ve been to some version of here before. So. Many. Times. Before. And we’re still here. People are still creating. Civilization has not (yet) collapsed under the weight of writing, printing, photography, radio, movies, typewriters, television, computers, the internet, or search engines. Each of them was supposed to be our undoing.

Let me guess. But it’s different this time? I’m sure it is. It was different all those other times too. It’s always already different this time. It’s whether the difference makes a difference that counts though. And so far, we’re still chugging along. (Train metaphor. Also failed to destroy us.)

But finally, the last thing is the simplest but also the hardest because it comes with real risk: Do what’s right for you. You know your work better than the online mob who would tell you what you can and can’t do. You cannot control what the mob will say or do. You can only control what you will do based on your own sense of craft and ethics related to your work. That doesn’t mean there can’t be negative consequences. But if there are, you can face them with a clean conscience at least. And in the end, that’s all any of us can do.

Thus, I’d encourage you to join me in taking stock of your life and your work, trusting yourself, and telling the world:

Clicking the link revealed the news. I was shocked, but not surprised.

This draconian state intervention did not emerge in a vacuum. It is an entirely foreseeable outcome from a discursive environment increasingly saturated by the convergence of AI hype and decades of cyber-doom rhetoric.

Subscribe now

The Fable of Mythos

I should start by being clear that the government’s actions are a clear case of overreach and were quite possibly undertaken for petty personal and corrupt financial motives. Nonetheless, Anthropic has spent weeks actively cultivating the very climate of fear that that the government has used to justify its actions.

In the weeks preceding the directive, Anthropic engaged in a highly publicized campaign of threat inflation regarding its Mythos model. In April, they claimed that it was so “strikingly capable” at uncovering and exploiting software vulnerabilities that it could not be released to the general public. Instead, they announced the launch of “Project Glasswing,” which restricted model access to a select consortium of corporations and government agencies.

News media dutifully played its part with The New York Times framing the moment as a watershed “reckoning”for global cybersecurity. The Economist proclaimed that Anthropic’s “warnings should not be dismissed.” The Guardian wins the prize, however, for taking it to the level of full-blown cyber-doom panic. A piece by Shakeel Hashim of Transformer.ai pointed to the 2024 Synnovis ransomware attack to warn about the potential dangers of Mythos. That attack, which he claimed “led to a patient’s death,” could be a harbinger of what is to come.[1]

Lethal cyber-attacks like this are thankfully rare. But a new AI release could change that – plunging us into a terrifying new world of chaos and disruption to the digital systems that we rely on. […] If such technology was widely available and as capable as Anthropic claims, the implications could be catastrophic.

He quoted “one cybersecurity expert” as posting on LinkedIn, “This is Y2K-level alarming.”

The only thing alarming is that anyone in 2026 can, with a straight face, still cite Y2K to warn of the dangers of impending cyber-doom.

Experts other than the one random guy-on-LinkedIn trolling for clicks soon pointed out that Anthropic’s warnings are mostly marketing hype. Gary Marcus wrote here on Substack that the fears about Mythos are “overblown” and concluded that

To a certain degree, I feel that we were played. The demo was definitely proof of concept that we need to get our regulatory and technical house in order, but not the immediate threat the media and public was lead to believe.

Georgetown University computer science professor, Cal Newport, had one of the best run-downs of situation on the “AI Reality Check” segment of his weekly podcast. He summarizes the various, independent third-party tests of Anthropic’s claims. They demonstrated that drastically cheaper, small open-weight models—some utilizing merely 3 billion parameters—successfully identified the exact same vulnerabilities as Mythos.

Responding to one such test, Yann LeCunn posted to X, “Mythos drama = BS from self-delusion.”

By marketing their model through the lens of catastrophic risk, Anthropic’s executives deployed fear to capture public attention while inflating the significance of their product. But, in doing so, they may have been too clever by half.

Subscribe now

When Technodoom Rhetorics Collide

Anthropic’s marketing strategy is, of course, exemplary of the insane tendency of frontier AI companies to use doom rhetoric to promote their products. Somehow, they’ve decided that telling people their products might take all the jobs and eventually kill us all is a winning strategy. This is not even the first time that Anthropic has used the possibility of cyber doom to market a new model. As Newport pointed out, they did the same with the release of Opus 4.6.

But it’s more than that. The marketing of Mythos (and Opus before that) is one more data point in the increasing convergence between AI hype narratives and cyber-doom rhetoric.

For over three decades, cyber-doom rhetoric has been a central feature of the U.S. public discourse about cybersecurity. Cyber‑doom rhetoric is a form of fear appeal that employs exaggerated, often fictional worst‑case scenarios alongside war and disaster metaphors to portray modern society and its economy as fragile, technology‑dependent, and always already on the brink of collapse. The goal is raising awareness and motivating policy or behavioral responses. Perhaps the quintessential example is the oft-warned about, but yet to arrive, “cyber Pearl Harbor”, where a massive cyberattack on critical infrastructure results in catastrophic physical, social, and economic collapse.

Not only are we are beginning to see this exact rhetorical playbook used in AI discourse, but we are also seeing the convergence of the AI hype/doom rhetoric with the preexisting cyber-doom rhetoric.

• Pearl Harbor metaphors? Check. A recent article in The Independent warned of an impending “AI Pearl Harbor,” imagining scenarios where malicious AI orchestrates massive blackouts or seizes critical infrastructure. See also the Shakeel Hashim article above, which is a textbook example of cyber-doom rhetoric (but without the direct Pearl Harbor analogy).

• Falsely attributing traditional cyberattacks to AI? Check, and check. A paper from MIT Sloan was withdrawn after falsely attributing 80% of ransomware attacks to AI. In another case, cybersecurity researchers heavily criticized Anthropic’s report about the supposed use of Claude Code to automate a Chinese cyber espionage campaign.

• Fictional scenarios meant to hype otherwise mundane events and turn our attention to cyber-doom? Check. We see this in a fictional narrative detailing rogue AI agents hacking municipal water treatment plants. At least one reader said they were “going to be sick” and had to “go sit down” because of how “chilling” this otherwise ridiculous piece of social science fiction was.

Even those usually skeptical of AI hype can sometimes be captured by the allure of cyber-doom rhetoric. In January 2025, for example, Gary Marcus warned, “2025 could easily see the largest cyberattack in history, taking down, at least for a little while, some sizeable piece of the world’s infrastructure.” “Four things,” he said, “could make it much worse in 2025.” Thing the first, second, and third were the same thing: generative AI, large language modes, and generative AI (again).

How did you survive the Great AI-Induced Cyber Pearl Harbor of 2025? Did you eat your leftover canned goods from Y2K or barter them for whiskey and cigarettes? Let me know in the comments below.

Share

Dangers of Doom

As my almost two decades of research on the use fear appeals and doom rhetoric in cybersecurity discourse has demonstrated, relying on metaphors of war, disaster, and existential threat to frame technological challenges can produce counterproductive policy outcomes. Such framings invite overreaction, anti-democratic decision-making, and an expansive license to centralized authority.

On Friday, Anthropic learned these lessons the hard way. Or, at least they should have. By leaning into cyber-doom rhetoric as a marketing strategy, Anthropic invited the state to treat its software as a weapon, ultimately encouraging the very centralized government overreach and suppression of democratic deliberation that critics of martial rhetoric have long warned about.

The use fear and doom rhetoric in AI discourse can also encourage fatalism and helplessness with respect to AI’s future. The supposed inevitability of catastrophic outcomes without the imposition of strict controls can discourage critical debate about alternative trajectories and responsible innovation. Stakeholders may adopt a defeatist stance, accepting AI apocalypse as unavoidable rather than exploring nuanced governance models or incremental safeguards that address the real, but less dramatic, risks of widespread AI adoption.

Conclusion

If it wasn’t clear before Friday evening, it should be clear now that we must find more productive ways to speak about AI, cybersecurity, and the intersections of the two. Policymaking must be grounded in empirical research regarding actual threats rather than hysterical, hypothetical scenarios generated by vendor marketing. That applies to AI companies and cybersecurity companies too, who also have an interest in inflating the threats they sell services to counter.

There are very real risks to widespread AI adoption, including cybersecurity risks. There are also very real benefits that we are only beginning to realize. But the risks and benefits include neither inevitable apocalypse, nor guaranteed abundance. We write the story of AI; it is not already written. We must reject the hype and doom from those who benefit from having us believe otherwise.

[1] While the cited source claims that “the attack contributed to the patient’s death,” it would be an exaggeration to say that the attack alone “led to a patient’s death.” Having read so many exaggerated accounts of cyberattack impacts that later turned out to be gross exaggerations, I would need to dig into the the details of the Synnovis incident before being sure that “contributed” is even accurate.

There is no shortage of advice available right now that purports to show how academics and other researchers and educators can use AI in their work. Most all of these takes are bad. They often come from people who have not spent enough time doing the work to know how to use AI responsibly or effectively. They hand too much (if not all) the work to AI. The latter is common when the advice giver is the former. In this post, I want to provide a glimpse of how I sometimes, but not always, integrate AI to augment the process of note taking I have been using for over three decades.

I’ll start by describing the note taking process that I use. Then, I will use a case study from my real work to demonstrate what this looks like in actual practice. Finally, I will offer some examples of how AI can assist in helping me with the mundane tasks of cleaning up those notes, as well as helping me zoom out and get a higher level view of the reading and note taking and commenting that I did manually.

The essential point here is that AI is used to assist and augment, not to replace. That includes helping with drudgery that is valuable but that I am unlikely to complete manually. It also includes providing new ways to query, connect, and “see” my work.

Subscribe now

Once A Debater, Always A Debater

I learned my current method of note taking in high school from a combination of my English and history teachers, as well as debate coach (who was also my freshman English teacher). In English and history, I learned to jot down facts, arguments, quotes, biographical details, and more on 4x6 note cards with one idea per card.

Similarly, on the debate team, we learned to take excerpts from source materials like books, journal or news articles, write a short “tagline” (paraphrase) for each, and tape them onto sheets of paper grouped by topic or argument. We called this “cutting cards” (see also HERE and HERE) even though we were using tape and paper, not note cards. The term harkened back to a bygone era in debate when students kept their evidence on note cards (like we were taught to do in English and history class).

What I did not know at the time is that these methods have deep historical roots in the Western tradition going all the the way back to the ancient Greeks. It is the method not just of the humanities but has been used by scientists too. It has gone by many different names over hundreds of years.

One of those names is “commonplacing.” This refers to a technique originally taught during Greek and Roman days as a memory aid for orators. One was instructed to construct a visualization of a known physical location and assign different information or parts of one’s speech to specific places within that location. Virtually moving from place to place in one’s mind would help the orator to recall essential information.

The method of categorizing information according to “common places” was adapted to the world of writing where scholars, rhetoricians, and those dirty sophists too, would take excerpts from reading and either cut and paste or copy them by hand under common “places” (i.e. “heads” or headings, topoi, loci) on loose sheets or in bound books. Those books became known as “commonplace books.”

Some form of this method has been used by scholars, especially in the humanities, ever since. In the Middle Ages, for example, one might have kept a “florilegium,” a gathering of “flowers” (e.g. excerpts and quotations). In modern times, the advent of the individual note card led to the emergence of the zettelkasten (“slip box”) method, which involved putting one idea per card and organizing the cards by topic. It’s the method I learned in school. It was made famous by the German sociologist Niklas Luhmann, though he certainly was not its inventor. I’ve written extensively on my personal website about the zettelkasen method, its merits, relationship to commonplacing, how I implement it, and how to get started with it. (See also the chapters on zettelkasten HERE.)

Share

“Cutting Cards” In The Wild

I am currently working on designing a new course for the upcoming fall semester. It’s a critical AI literacy course and in the early weeks I plan to offer a lecture and readings about the history of AI. Our assigned text for the course, Communicative AI: A Critical Introduction to Large Language Models, while good, does not offer as much on the history of AI as I would like.

So, I went on the hunt for article or chapter length treatments of the subject that I might assign in that week. I used a combination of traditional Google Scholar search with advanced operators, focusing on journals in the history of science and technology, Google Scholar AI search, and a Perplexity search. I filtered through the items myself, manually, skimming abstracts, looking at author credentials, and deciding if these were quality sources. If so, I clipped them to the Zotero collection I created for the course. (I’ve seen too many people lately say they are using AI to format their references. Don’t do that. It’s overkill and it doesn’t work reliably. There’s a free, dedicated tool for that task that works really well. Don’t reinvent the wheel with an LLM.)

I collected a lot more articles than just those covering history. So, I created sub-collections and started organizing the items based on the weeks in the course schedule and their lecture topics. (Lest you think I’m all clay tablets and papyrus over here, I spent a lot of time with Claude Code helping me sort out the schedule and topics. More on that in a future post. Rest assured, however, that was also an AI-in-the-loop process, standing by to assist rather than driving the train.)

So, I have a collection of sources about the history of AI. Now what?

I created an Andrey Karpathy-style LLM wiki, dumped them all into a folder in Obsidian and let Claude Code read them all write a wiki about it!

X NO! WRONG! X

You’re not “going deep” on anything if you’re letting the LLM do your reading for you! Later in the Gist, Karpathy talks about how he adds one article at a time, reading the AI summaries as they are produced, but adding that you could batch the process and have AI read a whole folder of sources for you and write the wiki.

Look, this is ingenious, no doubt. And it might work fine for whatever Karpathy is doing or whatever others are doing. But if you’re an academic researcher or educator, this is a disaster waiting to happen and you are a fraud if you rely on such a system.

As I tell my graduate students, “There’s no substitute for knowing WTF you’re talking about.” Deep subject matter expertise beats tools and methods. You only get that expertise by reading.

So, my next step was not a Karpathy wiki but reading and “cutting cards.” (<= Not X, but Y. I wrote that. Deal with it.) I did this using the excellent PDF reader built into Zotero. (If I’m reading articles clipped from the web, I use Readwise Reader. Ebooks, also Readwise. Paper books, I have a system for that too.)

I highlight relevant passages based on my current project or research questions. This is not comprehensive “coding” in the sense of qualitative methods. It’s looking for text that is responsive to my questions. Text that I feel meets that criteria—and it’s not an exact science (also this is a real human using an em dash, cry about it)—gets highlighted and I add a “tagline” (paraphrase) to the “comment” field for the highlight. If I have an actual comment, I add that to the text of the clipped passage itself at the bottom. I have a keyboard shortcut (“/c” that expands to **COMMENT:** and I add my thoughts. This renders as bolded markdown once it all gets exported to Obsidian.

There are a couple extra plugins in Zotero and Obsidian that make this work. In Zotero, that is Zotero Better Notes. In Obsidian, I use the Zotero Integration community plugin to import the full citation and “cards” for each source.

Again, the process is very similar for articles from the web, YouTube videos, podcast episodes, and ebooks. The first two go to Readwise Reader where I highlight and add a tagline. Podcasts get “highlighted” and excepts taken using Snipd, which feeds into Readwise. I read ebooks on either Kindle or (preferably) KOReader side loaded onto my Kobo device. Taglines are added as “notes” on highlights. They all get pulled into Obsidian using the Readwise plugin. (There are ways to do all of this that don’t involve Readwise too if you prefer to avoid it. Let me know if you’re interested in the comments and, if so, I’ll write up a post on how to do it.)

Subscribe now

AI Assistance & Augmentation

You might be saying, “We’re a long way into this post and you really haven’t used AI yet other than Google Scholar AI search and a Perplexity search.” Bingo! This is a human-first process. It is what I like to call “AI-in-the-loop” or “AI-standing-by.” (That rhymes, making me a poet, giving me even more humanities cred.)

I use AI to augment (not replace) a process I’ve honed over thirty years, with lots of experience and with a lot to show for it. That whole “judgement” and “taste” being “the moat” thing that everyone is on about? Yeah, this is where that comes from. Skip it and you’re a slippery slop slinger, which is worse than being a scruffy looking nerf herder.

Here are some ways I (sometimes) use AI to assist with my process:

“Coding” / “Tagging” the “Cards”

“Ah ha!” you say. “We’ve caught you using AI to do analysis for you!” But…nope.

In this case, AI assisted by helping me create the system I’d wanted inside Obsidian for years but couldn’t figure out how to make myself. So, instead, I had kludged together a system using to-do items and dataview queries that added any “card” with a to-do item mentioning another note to a dataview query that was dynamically populated in that other note. To-do items were the only way I could figure out for myself to reliably target the cards in the dataview query. (I know, I know. There’s probably another way. But I couldn’t figure it out and went with what I could get to work.)

There were several problems with this approach. First, the cards are not really “to-do items.” And I don’t need the false anxiety of a bunch of fake tasks just hanging around for no reason. Also, the system is not future proof. The related cards only show up dynamically when the page renders in preview mode. They do not live directly in the raw markdown document. In a future with no dataview, the connections disappear into the mists of digital history. Finally, too many dataview queries and/or too many responses to a query embedded in a note and Obsidian starts get sluggish. Again, these have to render ever single time you load the page.

So, I explained to AI what I really wanted and it walked me through how to set it up using a Templater template and script, both of which it created for me. Using this template/script along with Obsidian’s ability to “transclude” (i.e. embed) pieces of one note into another note, I achieve the trick of turning each item under a level 3 heading in a source note into an embedded card inside another note.

To code/tag a card, I run the Templater script, which provides a pop-up asking which other note I’d like to associate with the card. I search for what I need from a dynamically updating list, select, and hit enter. Then the level 3 heading of the card (which is its tagline) is pasted as a wiki link into the other note. Even if I no longer use Obsidian, the taglines and source note names will still live in those markdown files and I can find them again.

In this case, AI assisted by allowing me to create a simple system that automates copying and pasting across notes, which is otherwise a huge pain. That’s it. It’s not coding for me. Sorry haters. It’s just helping me do my work more easily and in a more future proof way.

Finding Related Notes

“There it is! He’s a fraud after all. He’s having AI pick the related notes for him!” Wrong again.

There’s a lot of talk about how “RAG is dead” lately. Traditional RAG has its limitations when you’re creating an “Agentic operating system” to replace yourself. But see, that’s not what we’re doing here. For what I’m doing, good old RAG is still very useful. And you should do what is useful for you in your work, not necessarily what some 25-year-old AI bro on the internet says is the new hotness.

Here are some examples of how I use semantic search to help me find connections between my notes that I may have overlooked.

Obsidian Copilot

I can set an entire note as context or select just card text inside a note, pass that to AI, and ask what other notes in the vault might be related. That can be done with either local or cloud models using the Copilot plugin. (The free version is all you need.)

Claude Code + QMD & Graphify

Claude Code uses a combination of QMD and Graphify to search the vault. QMD provides hybrid search, which is a combination of plane old keyword and semantic search together. Graphify creates a knowledge graph of the notes and allows AI to query the notes by traversing the connections between them. This tends to give more accurate results while consuming fewer tokens. It is more computationally expensive to create the graph in the first place and keep it updated, however.

I can use either cloud models—in this case I used Sonnet 4.6—or local models. (You can use Claude Code with local models alongside Ollama, just FYI.)

Mass Linking of Notes

Let’s cut to the chase. No, I’m not letting it just make whatever links it wants. Surely you’ve picked up the pattern by now, right?

One thing that I’ve wanted to do for a while is make sure I have notes for every person, organization, publication name, etc. I never get around to doing it, however, because it’s a lot of drudgery.

So in this example, I have asked Claude to enclose in square bracket wiki links all names of people found in the current note. The reason for doing this is that creating such links allows the creation of notes for those people with one click.

Once the note is created, Obsidian can itself find all linked and unliked mentions across the vault, allowing me to aggregate all content in the vault about a person, then read, compare and contrast, and write my own entry for them.

After the person’s page is created, Obsidian will show me all existing links to that page, as well as all mentions of the person’s name that are not directly linked yet. I can quickly click the “link” option under each unlinked mention in the right panel to create linked mentions. This automatically adds wiki links to all mentions of the person’s name in other notes without having to open those notes.

“Visualize” and “See” Notes

I will sometimes use AI to create high level summaries based on a list of specific notes given as context. In the example below, I created two summaries for Alan Turing based on all notes in my vault that mention him. One was made using a very small model running locally (Gemma4:e4b) and another using Sonnet 4.6 in Claude Code.

All AI generated content gets a prominent callout indicated that it is AI generated. Such content is meant to be a textual equivalent of the graph view, not text that counts as “writing” or that will go into a written product. It’s utilitarian, working text produced (not written) along the way.

Speaking of graph view, once I begin to link things, Obsidian’s own graph view can be helpful if used correctly. (Not the whole vault graph, which makes pretty pictures but is mostly useless, but the local graph per note. <= Not X, but Y…again! 😭)

Graphify also creates a human-readable graph in html that I can open in the browser.

Taken together, AI summaries and graph view provide another window into the notes in my vault, which now number over 2,000.

Connecting to and Creating with Other Tools

Claude Code (or Gemini CLI, or OpenCode, or Codex, or Pi, etc.) can use command line interface (CLI) or model context protocol (MCP) tools to connect what’s in the Obsidian vault to almost any other tool. Just to give examples of what’s possible, I prompted Claude Code to use the NotebookLM MCP tools to have NotebookLM create an overview report based on the notes I imported from Zotero.

I then instructed Claude to download the report and add it as a note inside of Obsidian. That is the note I used in the earlier screenshot demonstrating how to have AI automatically add wiki links to all person names.

Again, this is not text that I would use in my writing. And when Claude created a new note with this text, it added the usual callout at the top indicating it was AI generated, including that it was made by NotebookLM and the date created.

Finally, as an extra experiment, I had Claude instruct NotebookLM to create an infographic focused just on explaining the components of expert systems, a form of AI popular in the 1980s and 1990s that was discussed in some detail in one of the sources.

It’s an interesting result, though not one I’d use in this form. It’s too busy with entirely too much text that is hard to read. I don’t like the color scheme. But, then again, I didn’t give it much in the way of instruction other than “make an infographic about the components of expert systems.” With some refinement, however, this could be an image I use in a lecture slide.

Share

Possible Objections

“People use this stuff because they can’t do the work themselves!”

I’m sure some do. But you can look at my years of publications to see I know how to do the work without AI. I don’t need AI. Even though it helps find connections, for example, the vast majority of the links I make between my notes are still manual and I still tend to think up the more interesting connections among seemingly unrelated notes the old fashioned way. But not always. So I genuinely appreciate when AI suggests a connection I hadn’t considered, that prompts me to think, and ends up being something valuable.

I had a crude sort of system for this in the very first version of Evernote, which allowed attaching persistent Boolean and REGEX searches to tags. After adding a new note, I would have Evernote search the current note with all the tag searches to suggest tags I might want to add but that I had missed. It was extremely helpful during my dissertation research. That was just keyword and pattern matching. AI adds the semantic and graph search element, which is even more powerful.

“But you could write those summaries yourself! This is just the LLM wiki! How do you know it’s not all hallucinated slop?”

I can and probably will, too. But it’s still helpful to ask, “Hey what do I have in my vault?,” see a response, and save that response as a sort of work product like a log file. It’s a record of what was asked and answered if nothing else.

It’s not the LLM wiki because the AI summary is only created if I ask for one, not automatically. The summary is not the final word and is grounded in my paraphrases and comments which remain grounded in their associated excerpts.

That is how I know whether it’s hallucinated slop or not. Because I read and took notes on the material! I read and take the notes, then sometimes have AI help me get my bearings again, step back and see the forest for the trees before continuing on my way.

“But you could make that infographic or hire someone!”

No, I really couldn’t. I have neither the skill nor time to make graphics like that. I’m a one person operation. I have to get so many things prepared by the end of August that nice-to-haves like that infographic will fall to the bottom of the list and never happen.

And there is no universe in which I would hire someone to make such graphics. Course prep comes with a generous $0 budget. So, the ability that AI offers to create such graphics and other artifacts—e.g. interactive HTML elements that I will embed in my lecture slides and that will allow students to “play” the interactive elements as they follow along—provide an extra level of quality that I would not have been able to provide otherwise due to lack of time and money.

Conclusion

In this post, I’ve walked you through my human-first process of taking notes from source documents. That process is based in the zettelkasten and commonplacing tradition stretching back to the ancient Greeks. It’s been used by scholars in the humanities, scientists, and other intellectuals for more than a thousand years. It’s tried, true, and effective.

But I’ve also attempted to show how I sometimes use AI to augment what I already do by creating tools I couldn’t create otherwise, handling drudge work that still needs doing, finding and visualizing connections, and creating materials from my notes that I would otherwise not create.

None of the AI component replaces me. None of the AI component is even necessary. I don’t need it to write or to do my work. Too many influencers are promoting the creation of end-to-end agentic workflows or “operating systems” that hand their work over to AI. For academics, researchers, and educators, that is unacceptable and unwise. My goal is not even “human-in-the-loop,” but rather, “AI-in-the-loop,” “AI-standing-by” to assist with my work. There are so many valuable, little uses of AI that are genuinely helpful. Don’t sleep on those while pursuing a fool’s quest for AI to do it all.

I hope that someone will get some benefit out of what I’ve described here even if you never take up any of the AI elements that I’ve described. Ultimately, you need to reflect on how you work best, your own processes and values, and find the uses of AI that fit for you. Those might be small, simple things. Or it might be no AI use at all and that is OK too.

If you would like to speak more about how to do this in your own work, please reach out in a DM. I love to help people with note taking and research technology.

I know some of you will say the existence of “not x, but y” and em dashes or whatever your bugaboo is indicates this post is AI slop. After all, if you’re not completely rejecting AI, you must not be able to write and think on your own! So, if you believe in and care about such things, this is for you.

What Nano Banana thinks “Someone who is a skeptical enthusiast with respect to AI” looks like.

Introduction

In this first post at AI with Footnotes, I hope to provide a personal AI “manifesto” of sorts. My goal is to explain how and why I am approaching AI as what I call a “skeptical enthusiast.” I should probably start by addressing why it has taken me so long to publish my first post, however.

Go Slow, Start By Listening

You might notice that I created my account and this publication back in January 2026. I started by following all of the other writers on Substack that I could find who were writing on the kinds of issues in which I too am interested. When I did that, one of the first things that jumped out at me was the quantity of really good quality, nuanced thinking being produced by individuals from many different professions.

Of course, one of those is the Substack called Slow AI by Dr. Sam Illingworth. So, rather than just jump in and start putting my thoughts out into the world, I decided it might be better to take a slower approach, to get to know the community of AI writers on Substack, and to understand the themes and questions driving the conversation here.

That’s what I have done. Over the last three and a half months I have followed along, read and learned a lot from all of you. Now I think it’s time to start sharing some of my own thoughts. But first, I’d like to say more about why I started this project and how my background as an academic researcher and university professor shapes my thinking about artificial intelligence.

Polarized AI Discourse Makes Me Crazy

I come to the AI debate as someone who genuinely finds the technology fascinating and its potential impacts revolutionary, both good and bad. I think we are still in the earliest stages of this technology’s development and adoption. As such, I do not think any of us have all the answers.

But, I am surrounded by (over?) confident voices who say they’ve got it all figured out. AI is inherently evil and must be rejected outright! AI will usher in a new age of human flourishing and abundance! Human extinction is an acceptable price to pay for birthing superintelligent machines! (Um…WTAF on that last one?)

It all leaves me feeling like I’m taking crazy pills.

I came to Substack to set the record straight by shouting into the internet void: “It’s all more complicated! We need to slow down and learn! We need more nuance!”

What I found when I got here, however, is that I’m not alone. There are other people taking crazy pills too. One of the first posts I came across after signing up was one by one of my favorite writers here, one who doesn’t normally write about AI. In it, Charlotte en France lamented that, “The A.I. Revolution Was Stolen From Us.”

In her post, she described how she has no issue with AI in principle (same here), seeing its many potential benefits (again, same), but that she also recognizes that the future we are being sold by large corporations and billionaires is a lie (ditto). “I feel grief,” she writes, “not for the existence of A.I. but for a future surrendered to short-term profit and intellectual laziness.” Again, this is my feeling exactly.

Many others have expressed similar sentiments around the crazy-making lack of nuance combined with the grief of seeing an amazing new technology misappropriated from the start. In the comments to Charlotte’s post, Elaina Rowan wrote,

It is so rare to find a nuanced take like this when it comes to AI. People seem to either be super pro AI and think that everything AI is the best thing. Or they seem to be super anti-AI and say that every implementation of AI is the devil. And it makes me feel like a crazy person for having a take that is somewhere in the middle. So thank you for sharing your take. It is very valuable to have a nuanced voice in the matter.

More recently, a note posted by Kristi Pihl spoke to me. She wrote,

The hardest position to hold in the AI discourse right now isn’t skepticism or enthusiasm. It’s both at once.

The technology is genuinely extraordinary.

The damage done by the hype—by the investor class and non-technical C-suite treating it as a narrative prop instead of an engineering reality—is also real.

Nuance doesn’t travel well on the internet, but it’s the only honest place to stand. (Emphasis in original)

When I started this, I thought that there has to be more people like me, people who are not all hate or hype with respect to AI. What Sam, Charlotte, Elaina, Kristi, and so many others have taught me since I joined and started listening is that I’m not alone.

That’s who this publication is for.

• The other takers of crazy pills who are attempting to chart a path between the hype and the doom that dominate the AI discourse.

• Those who want more nuance and critical engagement instead of knee-jerk cheerleading or rejection.

• Those who do not see AI “going away” and believe, instead, that to shape its future in a positive direction we must learn all we can and bring that knowledge to the debate.

• Other professionals, especially researchers and educators of all kinds, who have real experience and expertise and who want to know how to mix that with AI, not to replace themselves or others, but to empower themselves and others to do more meaningful human work with less drudgery.

I call these people “skeptical enthusiasts.”

Skeptical Enthusiasm

The term “skeptical enthusiast,” though it may seem a contradiction, is meant to convey a productive tension. I think of the skeptical enthusiast as sharing a set of common characteristics.

Curious but cautious The skeptical enthusiast is genuinely curious about AI. This often involves experimenting with different models, following the latest in AI research, and imagining new uses for AI in their personal or professional life. At the same time, however, the skeptical enthusiast resists hype. Every bold claim triggers a quiet–or, let’s be honest, sometimes screaming–internal voice asking critical questions: “Compared to what?”; “Under which assumptions?”; “Using what data?”; “What was their research method?”; “Whose interest does this claim serve?”; etc.

Hands-on realism Skeptical enthusiasts don’t just read headlines. They dig into the details. They find the original study to make sure the results actually match the reporting. But they also get their hands dirty with the tools. They test tools for themselves. They say, “show me,” followed by, “hand me that thing so I can test it myself.” They want to know how and when AI systems work and, more importantly, where they fail.

Ethical and societal awareness While they are excited by AI’s potential, they remain deeply aware of and genuinely concerned about the many risks: bias, job loss, concentration of power, environmental cost, security and privacy. They worry that such an exciting and potentially beneficial new tool is not only being deployed on a mass scale without proper guardrails, but that lack of guardrails and regulation could undermine the technology itself. They want AI to succeed, but for the right reasons, at the right pace, and with the proper democratic oversight.

Comfortable with uncertainty Skeptical enthusiasts are comfortable saying, “I don’t know…yet.” Rather than jumping to definitive conclusions, they are okay saying: “This seems promising, but I have concerns and need to learn more.” They see skepticism not as negativity, but as critical discipline that restrains their enthusiasm. They see enthusiasm not as naive techno-optimism, but the motivation to dig deeper, to learn all they can and make better judgements.

Share

AI with Footnotes?

“AI with footnotes” has three meanings for me:

1. Going beyond the headlines.

2. “Yes, and…” “But also…”

3. AI in research and education.

I decided to start writing here because I think there is a need for a publication that goes beyond the hype and doom headlines and that analyzes both the good and the bad of AI. As an academic, I know the importance of digging into the details behind a claim. Some of the best material is hidden in the footnotes. I hope to bring that skill to my analysis of AI.

But also…in academic writing, a footnote, in addition to containing a source citation, often includes additional details, asides, qualifications or caveats, or “see also” material. AI with Footnotes aims to provide the same in the AI discourse. This is the “Yes, and” or “Yes, but” or “Maybe, but see also for more nuance,” etc.

I also hope to go beyond headlines by making this a space that doesn’t just report or comment on the bad and the good, but demonstrates it in practical terms. Thus, I plan to provide hands-on tutorials and demonstrations of what it looks like when AI works, but also when it goes wrong or doesn’t live up to the hype.

In short, my plan is to “learn in public,” sharing successes and failures, with the goal of learning how we can get more of the good and less of the bad out of AI.

How To Do It

What I am describing is an approach that I’ve taken to the other topics I have researched and written about over the years and comes from my training in science and technology studies. It’s what I think it means to “learn AI.”

For my dissertation and first book, which looked at the U.S. military’s use of nonlinear science–e.g. complexity theory, chaos theory, network theory–as a lens for understanding information-age warfare, foundational readings spanned military history, military theory, and all the major popular works on nonlinear science published to that point.

For my work on discourses of cyber conflict and security, I learned all I could about the technical aspects of cybersecurity and digital privacy. This included attending a two-day U.S. government workshop on critical infrastructure cybersecurity and, later, getting an MIT certificate in Cybersecurity for Critical Urban Infrastructure, along with many more hours over the years learning about the technical side of cybersecurity on my own.

Before I taught a course on drones at the University of Utah and published several essays about drone regulations, the First Amendment, and journalism, I bought and learned to fly hobbyist drones and even built one from scratch from parts I sourced from all over the internet.

I say all of this to say that I believe that to “learn” the thing and think deeply about it, you have to learn what you can about “the tech itself”–a problematic phrase as technology is always thoroughly social, of course–as well as the ethical, legal, social, political, economic, and cultural aspects. I think we position ourselves to do better at the latter by taking the time to do the former.

Conclusion

In short, for me, skeptical enthusiasm is a (sometimes uncomfortable) practice more than a definite position. It is my commitment to staying curious without becoming credulous, to engaging critically without retreating into knee-jerk rejection. If you have read this far and recognized yourself in any of this, then this publication is for you.

I do not claim to have all the answers about AI and I am skeptical of anyone else who thinks they they do. What I do have is an academic background in the history, philosophy, and sociology of science and technology, a genuine fascination with the technology, and a belief that working through hard questions in public—with appropriate footnotes—is more honest and ultimately more useful than performing certainty. That is what I will attempt to do here and I hope that you will join me.