On August 4, Google announced that it was providing K-12 students access to Gemini inside Google Classroom, including a “Learn with Gemini” option accessible directly from assignments in Classroom’s “Due soon” section. The announcement was quickly followed by the introduction of new tools across Google Search and Gemini, including study documents, interactive visuals, custom quizzes, and 3D simulations.

I am sure many educators will welcome these additions.

But now that AI has become directly accessible in school, it sparks two urgent questions:

What should AI do, once it’s embedded? And who gets to control it?

Now AI Is Contextually Aware

Standalone AI has been around for a while, so AI’s presence alone doesn't explain the current urgency. Part of the reason lies in AI’s rapid expansion from a text-based “answer machine” to a multimodal, conversational, tutor-like, and agentic performer. AI has not only expanded significantly into multimodal formats (image, audio, video, diagrams, etc.) and now provides live conversational interaction, but its AI-tutoring capabilities have also improved tremendously. Moving well beyond hints rather than answers and retrieval practice, AI tutoring systems are being designed to preserve productive struggle and promote desirable learning attributes. Meanwhile, agentic development has pushed beyond task planning and tool use into student workflows that include tackling student homework tasks and submitting completed work on their behalf.

But the most significant development is not that AI has become more powerful. It’s that AI has become more contextually aware and embedded where students already work.

A standalone AI chatbot has typically been detached from teacher intent. It knows very little about what a teacher is trying to accomplish and the actual conditions of student activity. As a result, its usefulness is limited. But Gemini in Google Classroom reduces that knowledge gap considerably. Gemini can see the assignments, handouts, and resources available to students in a particular class and use that information to provide relevant, focused support. And it’s positioned right where you would expect students to be looking for help: Assignment Due Soon.

With it comes a corresponding shift in student behavior and attitude. A separate AI chatbot visit is no longer needed; assistance is built in. There’s no reason for deliberate tool selection; convenience (and soon routine) predominate. AI is no longer banned; indeed, it’s actively encouraged. My outside-school habit can now become my inside-school one! I no longer have to feel guilty about using AI!

Tech companies have always tried to shape learning, but until Gen AI, their capabilities were noticeably limited. Edtech products were primarily algorithmic “decision-tree” systems with predefined rules and limited objectives. (And often subject-oriented: Here’s the math program. Follow the steps.) In contrast, generative AI can make open-ended decisions in real time for each individual learner—across any subject or topic— creating personalized instructional learning environments.

AI & Division of Tasks

AI uses probability and context-based semantic processes to analyze the learner’s reasoning, syntax, and specific knowledge gaps. While that’s impressive, it falls short of what a human can do. To start, a teacher draws on contextual knowledge and relational experience to “read” reluctance or frustration on a student’s face. A teacher can recognize when a student is overly tired, anxious, misunderstands a concept, or just needs a bit of encouragement. And they can make the necessary cognitive and psychological adjustments.

A teacher also sees and reacts to off-screen behaviour, unlike AI. They build a classroom culture and learn to read body language and tone. Their prior experiences and history with an individual student alert them when a student needs challenge, reassurance, space, or redirection.

One fundamental difference is that technology’s principal goal is often to move a student through a task. A teacher works to build the student’s ability to take charge of their own learning.

But it’s misleading to frame this relationship as “AI tutor versus human teacher.” A more accurate characterization is a division of tasks in which the AI can provide individualized support, but the teacher is responsible for analyzing AI activities within the broader context of the student learning process.

To wit, a recent review of 318 peer-reviewed publications on AI and teacher decision-making found human-AI collaboration to be the dominant narrative. The authors noted that teachers leaned into AI for learning analytics to identify student needs, but these could not replace the varied forms of student understanding and insight that teachers brought to the classroom. AI informed teachers, but humans made the final decisions.

A new SCALE report from Stanford University offers another way to define and evaluate this relationship. It uses a spectrum in which “high-intensity” is defined as a human tutor leading instruction with limited assistance from AI, and “low-intensity” as a student studying with an AI-only tutor without direct human oversight. As Stanford’s Lin Ler explains, the scale doesn’t categorize AI by sophistication; instead, it places it on a relational intensity scale. In other words, AI is defined by the depth and consistency of the human connection surrounding the student.

Perhaps unsurprisingly, the highest-impact tutoring occurred through live, human-led instruction. While they did not assume that more human contact would invariably make a tool better, the evidence points to limited AI effectiveness when the tool simply provides one-to-one responses.

To be clear, I don’t say this to diminish the significant progress Google is making toward pedagogically structured interaction. Their systems are increasingly better at preserving the “productive struggle” at the heart of learning. They ask probing, open-ended questions and encourage students to persist in solving problems. While Google is rather mum about how its products promote learning, it points to the impact of LearnLM, developed in conjunction with AI researchers, neuroscientists, and cognitive scientists. LearnLM was designed to enhance metacognition, active understanding, curiosity, and deeper exploration by promoting learning-linked behaviors.

Guided Learning builds on this progression by adding curriculum context and teacher oversight. For instance, teachers can assign a Guided Learning conversation tied to specific course materials and include it alongside the work students are engaged with. Teachers can receive individual and class-level information about student learning, and access chat transcripts. Overall, Gemini works within the context of the assignment in question and makes student learning visible to the teacher.

The “Platformarized” School

The researcher Luci Pangrazio writes about the “platformization” of schools, a broad shift in which school platforms don’t simply host educational activity; they increasingly influence it through their design and structure. As schools consolidate activities around one platform to improve efficiency, security, and interoperability and reduce maintenance, teachers lose some control over the tools and pedagogical approaches they use.

Generative AI may deepen this platformization beyond infrastructure that surrounds learning into instructional decisions that actively shape it.

As I’ve written, AI is becoming the operating layer of schoolwork and Google is leading the charge. Teachers already present with Slides, assign through Classroom, communicate through Gmail, and teach on Chromebooks, while students already write in Google Docs, store materials in Drive, search with Google, and watch videos through YouTube. So when Google introduces Gemini into Google Classroom, it is not just adding another element to a school technology repertoire.

It is changing the school’s work environment.

Conclusion

The question for schools is whether educators will redesign teaching, assessment, and student support enough to keep learning — not automation — at the center.

The unresolved tension is that teachers can oversee Google AI without actually controlling the pedagogy the AI executes. A teacher can decide to assign Guided Learning, assign course materials and provide learning instructions. But Google controls the model architecture, defines pedagogical behaviors, interprets productive struggle, sets scaffolding sequencing, shapes questioning patterns, and sets assistance thresholds. That’s a lot of instructional responsibility.

Schools need to look at Gemini through another lens. The baseline question of whether Gemini is safe, accurate, and compliant has not changed. But it needs to be accompanied by another question: Are we happy with the theory of learning that Google has embedded in Gemini—and, if not, what do we do about it?

At what point does the learning become so Google-designable that the teacher is no longer driving the bus?

Subscribe now

I’ve been preparing to lead two EdTechTeacher courses focused specifically on AI in international schools—one for school leaders and another for classroom educators—and the work has returned me to a professional environment that shaped a significant part of my career.

I taught history at an American school in France, an international school in Switzerland (that offered both American and British curricula), and in the American section of a French Lycée near Paris—where I assessed student work in accordance with French Ministry of Education guidelines. These experiences showed me that assessment traditions—whether American, British, or French—strongly shape what teachers, students, and parents define as “rigor.”

Later, during my years leading EdTechTeacher, I worked with international schools in Europe and Asia, presented at international schools conferences led by ELMLE and EARCOS, and served on a board charged with developing an IB curriculum for a STEM school in Saudi Arabia. From these experiences, I learned that international schools often strive to innovate but struggle with alignment—building cohesive practice in complex cultural environments.

Today, I live and work in Latin America. The societal conversation around AI in education here is considerably more upbeat than the one I encounter in the United States. A 2026 Digital Education Council survey found that 94% of higher-ed faculty polled in Latin America expect to use AI in their future teaching practices, and only 12% saw AI as a threat to their role. In contrast, US & Canada faculty intent to use AI has declined by 9 percentage points, from 76% in 2025 to 67% in 2026. While American AI-in-education conversations often center on concern and risk— cheating, cognitive offloading, robot teachers—in Latin America, discussion often begins with curiosity: “What is possible with AI?”

It’s one of several reminders that there is no single cultural response to AI.

And there is certainly no single “international school.”

An IB school in Switzerland, an American school in Italy, and a British international school in Singapore may all see themselves as international schools while functioning across diverse cultures, curricula, regulations, and expectations.

Yet many share an intriguing paradox when it comes to AI:

They may have more freedom to respond, but more complexity to navigate.

A Complex School Landscape

International schools often enjoy a degree of independence that many U.S. public schools do not. Schools in a public system often have to wait for district guidelines, state standards, approved app lists, or central-office policy before making significant changes. An international school can often move much faster. Like independent schools, they enjoy vast discretion over curriculum, assessment, and professional learning.

But an international school often asserts that autonomy across more layers of cultural, curricular, and institutional complexity.

An international school may be developing AI expectations for a faculty recruited from several national education systems, students working across multiple languages, and families bringing very different assumptions about technology, academic integrity, teacher authority, and what constitutes appropriate assistance.

It may also be operating within a host country’s regulatory environment while following an International Baccalaureate (IB), American, British, national, or hybrid curriculum. Its students may be preparing for universities in several different countries, each with varying (and evolving) expectations around AI use. Moreover, accrediting bodies, examination systems, boards, parents, and universities may all exert influence—and often in varied directions.

In March 2026, the IB acknowledged that its schools are navigating AI “under real pressure”: parents asking questions, teachers experimenting without clear guidelines, technology companies accelerating rapidly, and school leaders making important policy decisions with limited support. Significantly, the IB says its role is not to prescribe what schools must do but to provide principles that schools can adapt to their own contexts.

In all, freedom from one set of constraints can create responsibility for navigating many others.

When Educators Do Not Start From the Same Assumptions

Consider the faculty at the international school in Switzerland where I taught:

A half-dozen teachers were trained in the United States, about the same number in Britain, several others locally (or from other regions of Switzerland), a couple from France and Germany each. And one from Australia. (I was a hybrid; Canadian-American trained.)

We taught under different curricula, American or British. We used different assessment systems.

Now imagine us convening to discuss AI: We’d bring varied assumptions about authorship, student independence, appropriate assistance, and assessment.

Those differences might be manageable when teachers make individual classroom decisions, but they become difficult to manage when a school tries to establish shared expectations around AI.

Multilingual Learning Makes the AI Question Even More Interesting

Language adds another distinctly international dimension.

English is often the official language of an international school, but not necessarily the native language of its students.

Fortunately, AI can be extraordinarily helpful for multilingual learners. A student can receive explanations in their native language, practice speaking, get immediate feedback, or access material that might otherwise be linguistically inaccessible.

Empowering, certainly.

But AI can quietly do the language work the student was supposed to learn. And this is an acute issue facing international schools. It presents related dilemmas:

• When does writing assistance become substitution?

• When does translation become a scaffold?

• When does AI increase access, and when does it reduce productive struggle?

Stephanie Holt, Director of Learning and Teaching at DSB International School in Mumbai, explores this use of AI in an IB environment. DSB students use AI to create bilingual glossaries, rephrase instructions, and receive explanations at different levels of linguistic complexity while continuing to develop language proficiency.

In an autonomous school, different teachers may draw those lines in very different places—even for the same multilingual student.

So the challenge is not simply whether AI should be used. It is about deciding which forms of AI assistance advance learning goals and which bypass them.

Coherence: The Leadership Challenge

The globally recruited faculty member who prohibits AI may teach the same student as the teacher who actively encourages it. And the multilingual learner may receive AI support in one classroom and be penalized for essentially the same support in another.

Parents may get conflicting messages about whether AI represents opportunity, danger, cheating, tutoring, or essential preparation for the future.

Naturally, some variation is inevitable. Probably desirable. We shouldn’t expect (nor hope) that all faculty will respond in the same manner.

This can be particularly difficult in international communities precisely because the people involved do not necessarily begin with the same cultural or educational assumptions. Autonomy becomes complicated once again.

AI is creating pressing questions about assessment, academic integrity, student use, faculty expectations, professional learning, parent communication, and emerging technologies.

Diverse international communities make coherence harder. Rapidly changing AI makes coherence more urgent.

Coherence doesn’t mean forcing everyone to feel the same way about AI. It means replacing outdated assumptions with a shared, contemporary understanding of how learning happens in an AI-saturated world.

Building coherence across a globally recruited faculty, a multilingual student body, and competing international standards is daunting and frankly unattainable if leadership reacts to AI on a case-by-case basis. If an international school leadership team is committed to establishing shared expectations, they need to share a clear understanding of what AI actually is today—not what it was three years ago.

In that vein, here are four issues that I think all school leaders need to understand.

FALL 2026 COHORTS

Join me for two EdTechTeacher courses designed to help heads of school, curriculum leaders, technology directors, and other educational leaders make sense of a fast-changing AI landscape.

1. Schools are responding to an outdated version of AI

Naturally, the initial school response to generative AI focused on ChatGPT. All of a sudden, schools had to answer some fundamental, if vexing, questions: Can students use it? Should we block it? Did ChatGPT write this essay? Can we detect AI-generated writing?

At the time, those were understandable questions.

But “block or allow?” is not an adequate organizational framework for a school’s AI response. AI is simply changing too quickly, and becoming so deeply embedded in student lives,that schools can no longer organize their response to AI around the technology of 2023. But many school policies have been built around an outdated AI.

The IB has proclaimed repeatedly that banning AI is an ineffective response to innovation; its stated principles emphasize inquiry, educator agency, governance, transparency and continual adaptation.

To put in context, a trajectory of AI since 2023 looks something like this:

AI Chatbot → AI Multimodal Assistant → AI Tutor → AI Agent → Embedded AI

Today’s systems increasingly see, listen, speak, interpret images and video, create across media, conduct research, provide individualized feedback, and carry out multi-step tasks. Where is this considered in school AI policies?

We’re not witnessing the evolution of “a better chatbot.” We’re experiencing an AI that is changing the relationship between student and machine.

2. AI Is Becoming an Environment, Not a Destination

Consider this question for a moment:

If your school blocked every standalone generative-AI website today, how much AI would actually disappear from your students’ lives?

Not much, I’d venture. Because AI is there, in the background.

As I’ve written, AI is becoming a layer woven into the digital environments students already inhabit, such as Google Search, Google Docs, Gmail, Chrome, and YouTube. Not to mention learning platforms, research tools, and productivity suites.

Increasingly, students will not have to make a conscious decision to “go to AI.” It will just be there.

The Council of International Schools reports that many students do not always know when they are using AI because AI is increasingly embedded in ordinary online interactions.

For international schools—whose students may move across personal and school devices, multiple languages, national services, and different digital platforms—that boundary may be even harder to define.

3. Student Production Is Not Student Learning

AI is making student production dramatically more impressive. A student can produce a better essay, stronger slideshow, more sophisticated code, clearer explanation, or more polished research with AI than without it.

That is real production. But production and learning are not necessarily the same thing.

The OECD’s 2026 Digital Education Outlook notes that students using generic GenAI tools can produce better immediate responses while performing no better—or sometimes worse—when later assessed without AI. The OECD explicitly warns that AI can remove the productive struggle required for durable learning.

As I recently wrote, drawing in part on OECD’s Andreas Schleicher’s observations about AI and learning:

AI can increase the quality of what a student produces without necessarily increasing the quality of what a student understands.

A wonderfully written essay once provided at least some evidence about a student’s ability to research, organize, reason, and communicate. Today, it may provide considerably less.

This is especially important in IB and other inquiry-oriented environments, where the stated goals include reflection, agency, metacognition, and the development of the learner—not simply the production of an artifact.

The IB’s 2026 AI Design Principles explicitly call for AI that strengthens inquiry, critical thinking, creativity and collaborative problem-solving while preserving the “thinking, effort, and growth” central to IB learning.

In addition, the earlier-mentioned Mumbai IB case reports that assessment should focus on student judgment rather than polished output and describes students deciding what AI material to keep, adapt or reject.

If AI helps create a better product while bypassing the thinking the assignment was designed to develop, what exactly has been accomplished?

4. Schools Need Better Evidence of Learning

So, if polished production is no longer reliable evidence of student knowledge and understanding, what evidence of learning should schools value?

I explored this several years ago in Making Student Thinking Visible with AI, and more recently in writing about Stephen Fitzpatrick’s argument that AI has created a design crisis for schools.

If students can successfully complete assignments without engaging in the intellectual work those assignments were intended to develop, perhaps the challenge is not simply controlling AI.

Perhaps we need to reconsider what evidence we ask students to provide.

That might mean conversations. Oral defenses. Annotated sources. Drafts and revision histories. Reflection. Live problem solving. Students explaining where they used AI, what they rejected, and why.

The goal does not necessarily have to be making assignments “AI-proof.” That may be impossible—and ultimately beside the point.

Instead of focusing on Did AI do this?, we should increasingly be asking:

What evidence do we have that the student learned?

Conclusion

As I prepare for the international cohorts, I hope to help international school leaders understand the issues, examine school-based examples, and explore approaches to inform decision-making in their own contexts.

I don’t expect them to write a perfect AI policy.

But I want to ensure that their understanding of AI is aligned with the true state of AI. International schools have an unusual opportunity. Their autonomy can make them more flexible, experimental, and responsive than many other systems.

But their diversity also challenges them to create coherence across people who may begin with very different assumptions about learning, technology, culture, and AI.

Which makes deciding how to respond all the more important.

Subscribe now

Rebecca Winthrop is a senior fellow and director of the Center for Universal Education at the Brookings Institution. She focuses on global education, with a special focus on the skills young people need to thrive in work, life, and as constructive citizens. Michael Holt is co-founder of and a distinguished fellow at the Clayton Christensen Institute for Disruptive Innovation, a non-profit think tank, and he teaches at the Harvard Graduate School of Education. Andreas Schleicher is Director for Education and Skills at the Organization for Economic Co-operation and Development (OECD). He initiated and oversees the Program for International Student Assessment (PISA) and other international instruments.

Before delving in, let me make clear that this post is not a takedown. Far from it; I found the conversation thoughtful, substantive, and informative. It went well beyond the misguided “is AI good or bad?” question and focused on a deeper, more fundamental issue:

Does technology strengthen or weaken the intellectual and human work of education?

I mostly nodded as the presenters put forward their thoughts: AI should increase student thinking, support teacher agency, improve feedback, expand meaningful practice, and preserve human interaction.

But at times I found myself grimacing, especially when Winthrop and Schleicher agreed that general-purpose LLMs are “no good” when it comes to designing effective learning environments.

To start, I’ll focus on important areas of agreement:

AI has separated performance from learning

One of Schleicher’s most important claims is that generative AI has “divorced” task performance from learning. He references a group study in which 80% of students who used ChatGPT to write an essay could not recall any content from that essay. For Schleicher, this is a classic example of how AI can improve the quality of a performance without the students actually learning anything.

Schleicher then goes on to outline a contrasting scenario where students used AI to research a topic and explore different perspectives. This activity was undertaken before the students wrote the essay themselves. In this case, the students produced better work and learned more.

Schleicher’s conclusion is that the educational value of AI is not determined by its mere presence, but rather “the use of the tool and the design of the learning environments.” In other words: it comes down to how students use it.

Michael Horn reinforced Schleicher’s conclusion with a simple principle: “The person who does the work does the learning.”

The Importance of Pedagogical Intent

Schleicher repeatedly emphasizes the importance of pedagogical intent. He draws a distinction between creative uses of technology that increase cognitive activity and consumption-oriented uses that simply deliver information to students. Schleicher also contrasts teachers who actively design learning environments with teachers who become dependent on scripted programs and algorithms.

Schleicher cited Estonia as an example of a country that involves teachers directly in the research and design of AI-supported learning environments. In Estonia, teachers help determine what they want the technology to do rather than merely accepting tools created elsewhere.

As we know, the AI educational loop in other places is often different: This is the AI you’ll use. Here is your login. Here is your training. Now integrate it. In Estonia, teachers are not “in the loop.” They are busy co-designing it.

The Calligraphy Classroom: Instant AI Feedback

The conversation included only a few concrete examples of AI use, but one compelling example came out of a primary classroom in Shanghai.

Schleicher recounts watching students learning traditional Chinese calligraphy — a famously taxing undertaking. Students were forming the characters by hand, while scanners in their desks analyzed their work. Students received immediate feedback on the quality of their characters, and the teacher received information about where students were struggling.

Students were doing the work, but they were also receiving timely feedback. The teacher gains information about classwide needs. The brush never left the student’s hand, but AI shortened the gap between authentic student effort and meaningful feedback.

(I was hoping that Schleicher might mention that there are tools currently available to teachers that similarly provide students with immediate feedback and integrate teacher oversight. Snorkl, School AI Spaces, and Magic Student come to mind.)

AI Is Exposing Weaknesses Schools Already Had

Winthrop argues that AI has been layered onto an educational system superfocused on task completion, repetition, and the accumulation of knowledge (that students are not often asked to apply). She argues that AI literacy should focus on deep skills: understanding AI, thinking critically, making ethical judgments, creating, and solving problems. Schleicher provides a related argument that industrial-era education separated learning from assessment. Students spend years accumulating information and are then asked to reproduce it under highly constrained examination conditions. AI exposes the limitations of that model.

AI did not invent formulaic writing, superfluous homework, superficial research, and compliance-driven assignments.

But AI makes these weaknesses harder to ignore.

Cheating As the Solution

On a related note, Horn provocatively suggests that cheating may be “the solution, not the problem” because it exposes the shortcomings of a “transactional” education system. Horn argues that if an AI can readily pass a test or complete an assignment, the real flaw lies in the design of the test or assignment— not the AI itself.

Schleicher echoes that claim by describing cheating as a symptom of “poor education.” He points out that when learning is reduced to producing required results, students predictably search for the shortest route to those results. Compliance rather than learning becomes the objective.

To be clear, cheating is a problem. It damages trust between educators and students, misrepresents student understanding, and complicates the task of providing useful feedback.

But the AI cheating crisis surfaces a damaging assumption.

It points to where the educational system has confused the submission of work for the experience of learning.

AI-Aware, AI-Assisted, and AI-Resistant Learning

One of Winthrop’s most helpful constructs is that educators design learning that is “AI-aware, AI-assisted, and AI-resistant.”

According to Winthrop, AI-aware learning acknowledges that students have access to these systems. Educators avoid assigning work that can be trivially outsourced while pretending AI does not exist—and then relying on detectors to protect the assignment.

AI-assisted learning uses AI where it can meaningfully strengthen inquiry, practice, feedback, accessibility, differentiation, and revision.

AI-resistant learning preserves spaces in which students need to remember, think independently, communicate orally, work directly with other people, and navigate the friction of human relationships.

One of its strengths is that it stays clear of universal declarations that AI should either be allowed everywhere or banned everywhere.

A question for educators then becomes:

Which parts of a particular learning experience should be AI-aware, AI-assisted, or AI-resistant?

Where I disagree

Little good from general-purpose AI tools

Schleicher says he has seen little good emerge from general-purpose AI tools and claims that most strong educational applications are specifically designed for pedagogical purposes.

Winthrop similarly favors “narrow” targeted educational AI tools over “wide” or broad ones because she has understandable concerns about how young people are engaging with general-purpose chatbots.

I say “understandable” because general-purpose AI can hallucinate, write work for students, expose students to inappropriate interactions, and encourage AI dependence.

But “general-purpose” does not mean “educationally purposeless.”

I wrote recently on LinkedIn that while it’s true that narrow, highly specialized AI systems can be aligned to a particular curriculum and constrained by certain safeguards, that does not mean that general-purpose LLMs are fundamentally less educational.

What matters is whether the AI in question is leveraged effectively through carefully constructed instructions, clearly defined roles, and clear expectations for student thinking.

ChatGPT, for example, is wonderful AI for language learning as it can serve many useful roles. A language teacher can configure ChatGPT Voice to be an excellent conversation partner that asks appropriate follow-up questions, focuses on specific errors, and adjusts its phrasing to match the student’s ability.

General-purpose LLMs offer more flexibility than specialized AI— but less control. They can address support various stages of the learning process and assume different roles—depending on the teacher’s objective—such as: a historical figure in a source-grounded simulation, a debate opponent, a skeptical reader, an interviewer, a revision assistant, and a metacognitive reflection partner.

But “wide” AI can elicit a “wide” range of responses, especially if prompting is weak and inconsistent. It requires stronger instructional guidance, which calls for a deeper understanding of general-purpose LLMs and how to maximize their effectiveness.

But a trade-off is not proof that “narrow” is better than “wide.”

The key question, however, is not specialized versus general-purpose AI. The best instructional strategies are likely to incorporate both. The question is whether AI usage is carefully and purposefully designed, rather than educationally rudderless.

“Avoid tool teaching”

Winthrop advises against teaching students or educators the functions of individual tools because tools change rapidly. She argues that AI literacy should focus on deeper capabilities: understanding AI, thinking critically, making ethical judgments, creating, and solving problems.

Her larger point is on target, but she overlooks how teachers typically become immersed in technology. For sure, school PD should not consist of an extended presentation of prompts and tool features. But the recommendation to avoid tool teaching entirely is misguided.

For many teachers, the tool is the doorway. Many teachers enter the world of AI because they need it to accomplish a particular task. For instance: generate examples for a math lesson, develop reading questions, improve feedback on student writing, practice speaking for students, and adjust the reading level of a primary source.

That immediate, practical need creates curiosity and willingness to experiment—states of mind needed for effective exploration.

Yes, the tool is the hook. But it’s not the whole lesson.

When teachers use a tool, they need to learn more than which button to press. They need to understand the purpose, the limitations, and what good output looks like.

“Most powerful uses” are learning analytics

Schleicher argues that the most powerful uses of AI are not directly student-facing. He points to learning analytics that help teachers understand student progress and schools improve their practices.

There is certainly much to applaud. Teacher-facing AI can help educators: adapt materials, identify student learning patters, provide more precise feedback, create differentiated supports, and develop better targeted instruction.

Yet, teacher-facing AI should never be the final destination.

Students ultimately need active, guided experience using AI. They need to learn how to use it to explore ideas, to question it, challenge it, and verify it.

Bottom line: We cannot prepare students for an AI-infused world by keeping AI exclusively on the teacher’s side of the desk.

Amplifier — For better or worse

The Brookings conversation is at its best when the speakers are discussing whether technology increases or diminishes student thinking, teacher agency, meaningful feedback, and human interaction.

They correctly point out that AI amplifies what already exists. Schleicher, in particular, characterizes AI as an accelerator and amplifier of both good and poor educational practice.

As for tools, schools do not need more technology or less technology. They need better educational judgment. It doesn’t so much matter if the AI is “wide” or “narrow” if it’s leveraged purposefully and carefully. The AI also need not be built specifically for education. What matters is that students are building knowledge, exercising judgment, receiving meaningful feedback, and undertaking the intellectual tasks that learning requires.

For many educators, that fact alone is enough to end the conversation.

We all know that schools have spent the past several years attempting to stop students from palming off AI-generated writing as their own. AI has prompted teachers to rewrite assignments, scrutinize revision histories, and otherwise search for signs that a submitted essay is not actually the student’s work.

And then there is Douglas. An experienced English teacher, computer science teacher, and AI literacy leader, Douglas willingly helps students defeat the very systems designed to preserve academic integrity.

So, when I interviewed him recently, I started with the obvious question:

“What were you thinking?!”

His answer was a lot more thoughtful than my grab-your-attention headline implies.

This is my favorite interview to date. I hope you’ll watch it.

Subscribe now

More Than a Provocation

To be clear, Douglas did not ask students to generate an essay, disguise it, and submit it dishonestly. Indeed, he asked them to study their own writing first by asking the following questions:

• What does your voice sound like?

• What words, sentence patterns, rhythms, and habits make your writing recognizably yours?

• What does AI-generated prose tend to do that you would never naturally do?

Students compared their own work with AI-generated text, identified the differences, and developed instructions that could help AI produce language closer to their actual voice.

Okay, you still might want to end the conversation.

Many teachers would understandably argue that the process of writing sentences is part of the thinking process itself. We do not always form ideas first and then simply convert them into words. Often, we uncover what we think through the struggle to write. From that perspective, revising AI-generated prose is not the same as writing.

But Douglas’s experiment forces us to address a difficult question:

A Classroom Laboratory for the Rest of Us

Douglas teaches at Menlo School, an independent school in Silicon Valley. As you might imagine, he works in an environment with strong technology access. He also benefits from a supportive administration, parents working in technology, and enough instructional freedom to experiment. His school also offers a freshman course in AI literacy and fluency. Douglas can redesign assessments, invite candid discussions about AI use, ask parents to contribute, and implement instructional practices that would trigger immediate resistance in most other school systems.

But that does not make Douglas’s teaching irrelevant.

If anything, it makes his classroom a laboratory for the rest of us.

Douglas’s laboratory does not aim to create a model everyone must follow. Rather, it is to test assumptions, uncover problems, probe solutions, and provide information to evaluate. Douglas can certainly move farther than most educators, but the questions that come from his classroom will eventually reach every school.

And the most provocative of these concerns writing itself.

When Writing Becomes Directing

Douglas argues that writing has acquired another layer. Traditionally, writing instruction has strictly emphasized close writing. The writer chooses words carefully, building sentences, shaping paragraphs, organizing arguments, and refining style. Douglas is not saying these skills are disappearing. But he believes they are the foundation for something else: writing as directing.

He uses the analogy of a film director. A director does not personally operate every camera, design every set, compose every scene, or deliver every line. The director establishes the vision, evaluates options, makes decisions, and is responsible for the final product. Douglas believes writers are increasingly doing something similar with AI. They develop the main idea, organize a structure, define a tone, direct different systems to produce drafts, compare the results, and assemble a final piece that reflects their intent.

Of course, for writing teachers, this comes across as supervision rather than authorship.

Douglas’s retort is that good directing requires deep knowledge of the craft. A person who cannot recognize a weak argument, a generic sentence, a misleading claim, or a stylistic mismatch will not direct AI well. He argues that close reading and close writing do not become less important. They become the foundation for making sound judgments about AI-generated language.

And many in our society are already attempting to direct AI.

The problem is that students haven’t developed the knowledge and judgment required to do so effectively. A student who has never struggled to construct an argument has no foundation for evaluating one. A student who has never developed their voice when writing will not know if AI has captured it. A student who lacks subject knowledge may accept polished nonsense because it sounds convincing.

Douglas’s classroom points toward a future in which schools may need to teach both independent composition and AI-supported direction. I’m sure that’s bound to unsettle most educators, but similar changes are already occurring in professional work. Software engineers supervise AI-generated code. Marketers compare machine-produced publicity. Lawyers review AI-generated contract drafts. In this environment, the human element revolves around judgment, context, and responsibility.

The Skills AI Cannot Replace

Douglas’s other ideas reinforce these human elements.

He warns that “accuracy breeds complacency.” Educators often focus on what LLMs get wrong, but Douglas argues that AI’s greater danger may be that it is most often right. As AI performs more reliably, people stop checking its output and gradually let go of the measures required to identify the occasional serious error. The challenge becomes deciding which abilities students must continue to practice even when AI can perform them accurately.

Douglas also talks about the peril of “unseen unseen,” the ideas or perspectives that AI fails to provide us. When a chatbot produces an answer, it limits the field of possibilities because it’s always presenting the next most-probable step. Students instinctively react to what appears on the screen and often don’t consider what is missing. As a result, Douglas asks students to brainstorm before using AI. He doesn’t do so because he’s sure the student’s first idea will be better, but because AI should not limit the possibilities before the student has entered its space.

Douglas also draws an important distinction between AI literacy and AI fluency. He points out that teachers are often thoughtful about when AI should or should not be used, but many lack practical “in the weeds” experience with the tools. Students, in contrast, use AI constantly, but fail to reflect on why they are using it or how it might impact their learning. In other words, teachers and students possess opposite halves of the competence schools need.

It may surprise you that Douglas does not saturate his classroom with AI. In fact, he has brought back more pencils, paper, quizzes, and in-class checks. Students may use AI on take-home assignments, but they must also demonstrate understanding without it. In this way, he separates supported practice from independent proof. This may be one of the most practical lessons teachers can take from his work.

Students Need Teachers More, Not Less

Douglas’s progressive activities may not travel to many other classrooms. But the underlying questions certainly do.

• What should students be able to do without AI?

• When should human thinking come first?

• How should students explain their use of AI?

• What evidence shows that learning actually occurred?

• Which skills must be protected even when technology can perform them faster?

Which brings us to Douglas’s most important argument:

Students need teachers more, not less, in the age of AI.

Teachers have legitimate concerns about privacy, bias, environmental cost, misinformation, dishonesty, and cognitive offloading. Douglas does not dismiss them. He argues that these concerns make teacher involvement more necessary. Students are already using AI. If teachers refuse to engage, it does not create an AI-free world. It may leave young people to navigate powerful systems without adults who understand learning, ethics, and responsibility.

AI can produce answers, explanations, examples, drafts, and feedback in seconds. It cannot decide what a student should value, what struggle is worth preserving, or whether efficient completion has replaced learning.

A Path Forward, Not a Blueprint

Douglas’s teaching may seem radical. His ideas are much harder to dismiss. He is not offering a universal blueprint, and not every experiment should be copied. But his classroom gives us an early look at the choices schools will soon have to make.

The greatest danger is not that teachers will move too quickly.

It’s that they will refuse to move at all, and leave students to find their way alone.

They are fully aware that AI can solve a problem, show the steps, explain the method, check an answer, and produce a polished explanation in seconds. For teachers who have spent years trying to help students build conceptual understanding, that raises an obvious concern: What happens when students can bypass the thinking?

Karen Levin hears that concern all the time.

Karen, founder of Math for Humans, brings an unusual blend of experience to the AI conversation. She began as a high school math teacher in Boston, worked as a math coach, supported blended learning initiatives across schools, and now helps educators think about math teaching in an age of AI.

When she works with math teachers, she often begins by asking them what worries them most. The word that rises to the top is almost always cheating.

But cheating is only part of the challenge.

Teachers also worry about inaccuracy. They have heard that AI can solve advanced Olympiad-level math problems, but they have also seen it make mistakes on basic arithmetic. They worry about students outsourcing their thinking. They worry that AI will become a crutch. They worry that students will get the answer without developing the understanding.

You get the picture.

But Karen’s key contribution is that she reframes the question. The issue is not simply whether students can get an answer from AI. The deeper question is this:

How does the tool position students as mathematical thinkers?

That may be the most important AI question for math classrooms right now.

Listen to my conversation with Karen Levin

The Answer Was Never the Whole Story

Math learning cannot be reduced to answer production. A correct answer can hide shallow understanding. An incorrect answer can reveal thoughtful reasoning. A student may understand the concept but struggle with notation. Another may know the procedure but be clueless as to why it works. A third may be able to solve the problem silently but be unable to explain the reasoning behind it.

In math, the evidence of learning often lives in the messy middle, such as the strategy attempted, the explanation offered, the error noticed, the revision made.

AI did not invent answer-getting. It merely accelerates it.

Karen made a point in our conversation that surprised me. She said she always liked that the answers were in the back of the textbook. Not because the answer did not matter, but because the answer was never the whole story. What mattered was how students were thinking about the problem, what method they were using, and what they understood about the mathematical idea.

Boy, that distinction matters even more now.

If AI turns math into a faster path from problem to answer, then we have a problem. But if AI can help students reason, explain, question, compare, and revise, then we may have something worth exploring.

Step-Giver or Sense-Maker?

The difference is between AI as a step-giver and AI as a sense-maker.

The step-giver says: Here is step one, here is step two, here is step three.

The sense-maker asks: Show me what you tried. Why did you choose that strategy? Where are you uncertain? Can you explain why that step works? What would happen if the numbers changed?

The sense-maker is instructional.

A step-giver may help a student complete a problem. Sometimes that may even be useful. Students need models. At times, they need hints. They need worked examples, too. They also need support when they are stuck.

But if the dominant AI experience is “give me the steps so I can follow them,” then students may learn to imitate mathematical procedures without developing mathematical judgment.

A sense-making tool, by contrast, keeps the student inside the thinking. It does not treat the student as a passive recipient of explanation. It asks the student to contribute, explain, justify, and revise.

Making Mathematical Thinking Visible

That is where Karen’s frame connects so strongly to the underlying challenge of making student thinking visible.

In a world of AI-generated answers, teachers need more than final products. They need evidence of process. They need to see how students are approaching the work, where their reasoning is developing, and where misconceptions are taking shape.

In math, that evidence might take many forms. A drawing. A spoken explanation. A comparison of two solution paths. A reflection on where the student got stuck. An error analysis. A revised explanation after feedback.

This is why tools that ask students to explain their reasoning are so interesting.

In our conversation, we discussed Snorkl, a platform that allows students to draw, write, and record their voice as they explain their mathematical thinking. Teachers can create criteria, and the AI can provide feedback aligned to those criteria. A student might get the answer right but still be asked to explain more clearly, justify a step, or show deeper understanding.

That is a very different use of AI from simply asking a chatbot to solve the problem.

It also reveals a possibility that is often overlooked in conversations about AI and math: maybe one of AI’s most valuable uses in math is not solving, but speaking.

Karen emphasized the importance of math discourse. When she coaches teachers, she notices the familiar classroom pattern: teacher asks, student answers, teacher responds, another student answers, teacher responds. The teacher remains the center of the conversation.

But strong math classrooms require something more. Students need to talk to one another. They need to hear different strategies. They need to explain their thinking out loud, defend it, question it, and refine it.

That kind of discourse is hard to build.

It is especially hard for multilingual learners, who may be developing mathematical understanding and academic language at the same time. But it is hard for native English speakers, too. Explaining math clearly is difficult because it requires more than getting the answer. It requires organizing thought.

In a past post I related the experiences of Bret Xu, a NYC math teacher whose students use Snorkl to explain their mathematical reasoning aloud. Many of them are English language learners. So when they explain a math problem, they are practicing both mathematical reasoning and English academic language.

And they groan when he tells them they are going to use it.

That may sound like failure. But actually it’s the opposite.

They groan because it is hard. They have to think. They have to explain. They have to make their reasoning visible. They cannot simply produce an answer and move on.

That is productive friction.

And productive friction is precisely what AI can easily remove if we are not careful.

AI Tutoring Needs Architecture

This brings us to the crutch problem.

Bret built a custom AI tutor for his students, many of whom arrive with wide differences in math preparation and language background. The tool can respond in Spanish when needed. It can support students at different levels. It can extend his reach as a teacher.

That is powerful.

But he also noticed that students sometimes go to the tutor too quickly. They use it before they have tried to develop a strategy themselves. The tool is so helpful that it can become a shortcut around productive struggle.

This is not an argument against AI tutoring. It is an argument for better architecture.

Karen is cautious here, and I think rightly so. She is not calling for every school to rush into broad student-facing AI use in math. She sees real promise, especially with teachers who are early adopters and who keep student thinking at the center. But she also notes that we do not yet have enough evidence to assume that whole-school AI use with students in math will go well.

That caution is worth noting.

It also points to a practical starting place: teacher use.

Karen sees strong possibilities for teachers using AI to improve planning, differentiation, discourse, language access, and curriculum-connected lesson design. But she is not talking about vague prompts like “differentiate this lesson.” She argues for context. Give the AI research-based frameworks. Connect it to the curriculum. Use it to solve a specific instructional problem.

That echoes lessons from earlier edtech and blended learning efforts. Technology works better when it is connected to a clear learning goal, embedded in the curriculum, used with fidelity, supported by teacher training, and given time inside the school day.

AI is not exempt from those conditions.

In fact, it may need them even more.

So perhaps math departments should begin not with a yes-or-no debate about AI access, but with a better set of questions.

When students use AI, what evidence of mathematical thinking remains?

Did the tool help them reason, explain, revise, compare, question, and justify?

Did it keep them inside the productive struggle long enough to learn?

Or did it simply move them more efficiently from problem to answer?

Karen Levin’s challenge is useful because it shifts the conversation from access to design. AI tools will keep getting better at solving math problems. That is not the hardest question anymore.

The harder and more important task is making sure students keep getting better at thinking mathematically.

Many are writing policies, holding faculty meetings, warning students, testing tools, debating detection, and worrying—quite understandably—about cheating and cognitive offloading.

But many of these conversations remain focused on control—what to allow, block, or detect—when the deeper challenge is design. In an age of agentic AI, schools must redesign learning for a world where AI can manage the entire student workflow, not just provide answers.

Across disciplines, immediate concerns often differ. English and history teachers worry about AI-generated writing. Math and science teachers worry about solved problem sets and bypassed thinking. World language teachers worry about translation and conversational AI. But underneath these disciplinary differences is the same fundamental design challenge: AI is moving faster than our assignments, assessments, and school policies.

The question is no longer simply whether students used AI. It is what kind of thinking the assignment was meant to develop, what evidence of that thinking remains visible, and where AI support crosses the line into cognitive offloading.

Subscribe now

A Practitioner’s View from the Messy Middle

These challenges formed the core of my recent conversation with Stephen Fitzpatrick. Stephen is a veteran history teacher, debate coach, parent, and former curriculum director who has spent more than three decades in the classroom. He teaches American history, government, independent research, and “great books” at an independent K–12 school north of New York City. He also advises an AI club, writes the thoughtful Substack Teaching in the Age of AI, and works with schools trying to make sense of this moment.

Because of this background, he doesn’t look at technology through the lens of industry hype. He looks at it from the perspective of daily classroom realities, where genuine educational progress is either realized or not.

Watch our conversation on YouTube

When Stephen first saw ChatGPT in late 2022, he knew immediately that it was “going to be a big deal.” He started reading, testing, experimenting, and running his own assignments through AI systems to see what they could do. But the question he kept returning to was not “What else can ChatGPT do?” It was the more important classroom question:

“How does this work in teaching and learning?”

That question has become harder to answer.

Students Are Already Living in the AI Layer

One of Stephen’s most important observations is that students are using AI more often, and in more varied ways, than many teachers realize.

And not just to write papers.

Students are using AI to organize notes, generate practice questions, create study guides, prepare for tests, revise writing, navigate difficult math problems, and produce visual explanations. In Stephen’s AI club, students are candid about their use of AI (and that of their classmates) because the conversation is not a disciplinary investigation. The stories that surface go well beyond the caricature of the student who pastes a prompt into ChatGPT, receives an essay, and submits it as original work.

That happens, yes, but students are leveraging AI in diverse ways. Some are trying to use AI as a tutor. Some are using it as a study partner. Some are using it as an organizer, explainer, quiz generator, or research assistant. In all, some use it thoughtfully. Some offload too much. And many are unsure where the boundaries are.

That uncertainty is one of the most important themes of our conversation.

The Assessment Continuum

As Stephen put it, policies can sound clear in the abstract but become much harder to apply inside actual classrooms. For instance, we all agree that a student should not ask AI to write an assignment and submit the output as their own. But what about AI for brainstorming ideas? Building a study plan? Grammar checking? Using AI to locate sources? Summarizing a research study? Asking for feedback on a thesis statement?

The middle gets mushy awfully quick, and that’s where the real work is.

Stephen tells students something revealing when they ask whether they can use AI to for a particular task. First, he asks, in effect: if I said no, what would actually stop you? Then he asks the more important question: does it help you learn and perform?

That second question is everything.

One student told Stephen he used AI to prepare for a chemistry midterm and then performed poorly. Stephen’s response was basically: there’s your feedback. In other words, the outcome itself was the most valuable feedback available. The tool had failed to advance actual comprehension, and the test score simply exposed that gap.

That is where schools need to go next. Not toward blanket permission. Not toward blanket prohibition. Toward sharper judgments about learning in context.

You Cannot Ban What Has Become Ambient

This is where many school AI policies start to wobble.

A policy that targets ChatGPT but ignores Google Search, Google Docs, YouTube, Google Classroom, databases, writing assistants, and embedded AI dashboards is not an AI policy. It is a time capsule.

We all know a teacher telling students not to use ChatGPT.

But what to tell students now, when Google Search can summarize? When YouTube can answer questions? When Docs can take half-formed thoughts and turn them into essays ? When Slides can create a slideshow from a one-sentence directive? When Classroom-adjacent workflows generate, organize, and suggest? When an AI assistant moves across tools rather than sits inside a single chat window?

The more AI becomes ambient, the less useful it is to define academic integrity by tool usage alone. Schools need principles that travel across tools, contexts, and assignments.

Stephen makes this point bluntly. When people say “ban AI,” he wants to know what they actually mean. A teacher may decide that students cannot use AI for a particular assignment. That is one thing. But a school claiming to ban AI at large is increasingly like a school claiming to ban the internet. AI is entering search, productivity tools, learning platforms, data systems, and the ordinary interfaces students and teachers already use.

This is a shift in the work environment.

Stephen illustrated this by highlighting a telling example from outside the classroom. He began discussing Veracross, the enterprise management system used by many independent schools. Veracross has integrated cloud-based AI across its entire platform, allowing administrators to instantly build custom dashboards and synthesize complex data spanning admissions, advancement, and academic programs.

Herein lies a striking institutional paradox. While faculty rooms remain locked in debates over blocking individual text-generation tools, the administrative and business operations of schools are sprinting into an AI-driven future.

If students experience a school that uses cutting-edge, predictive AI to manage their admissions, enrollment, and sports schedules, but walks into a classroom focused on 2010-era "compliance and lockdown" tactics, does not the school appear profoundly disconnected and the classroom out of touch?

As I wrote regarding Google I/O 2026, teachers already write in Google Docs, present with Slides, assign through Classroom, search with Google, show videos through YouTube, and teach on Chromebooks. When Google announces changes in the AI capabilities around those tools, it is not simply adding another element to the edtech shelf. It is changing the operating layer of school work.

Google is now the 800-pound elephant in the classroom.

Artificial intelligence does not require an invitation from school leadership to impact the classroom. Instead, it has already been integrated directly into the core platforms that dictate daily workflows—altering how teachers plan, communicate, search, and provide feedback — and how students research, plan, organize, communicate, and prepare.

The Policy Gap Is Really a Design Gap

A gap Stephen identifies is that school AI policies often try to solve a design challenge with compliance language.

A school can say, “Do not use AI unless approved.” A teacher can say, “AI is prohibited on this assignment.” A syllabus can list permitted and prohibited uses. Those moves may be necessary, but they are not enough. They do not answer the more fundamental instructional question:

What cognitive work is this assignment supposed to develop, and how will we know whether the student actually did it?

AI has exposed something that was already shaky in many school assignments. For years, teachers have often relied on the final product as the main evidence of learning: the essay, the lab report, the problem set, the project. But when AI can produce polished final products, the final product becomes less reliable as a sole measure of student thinking.

AI did not break school assignments. It revealed what many assignments were failing to measure.

Stephen makes this point in a practical way. AI, he says, forces teachers to break down their pedagogy into more discrete steps. We may think we already do this, but the reality of teaching is often messier. We hand out assignments without always clarifying which parts of the task matter most. Is the goal to gather information? Evaluate sources? Build an argument? Practice writing? Demonstrate recall? Show creativity? Synthesize multiple perspectives? Revise one’s thinking?

If we are not clear, AI will make the confusion apparent.

That is why the most important AI conversation is “What are we asking students to learn?”

The Ghost of EdTech Past

Stephen’s perspective is valuable precisely because it is devoid of tech-utopian hype. In fact, he admits that his initial, pure optimism during the first several months of the ChatGPT era was eventually met with a “big splash of cold water”.

That cold water came in the form of reading Failure to Disrupt, a book by MIT’s Justin Reich that meticulously chronicles how educational technology has historically overpromised and underdelivered. Reich’s main argument is that educational technology consistently fails to disrupt because it adapts to old school structures rather than changing them, ultimately reinforcing the very systems it promised to transform. I couldn’t help but smile inside when Stephen brought this up, as I had discussed Failure to Disrupt with Justin several months before it was published. Back when Justin and I were just two history teachers sharing a classroom at an independent school in Boston—and later co-led EdTechTeacher—we spent hours analyzing shortcomings in educational technology.

In our last conversation, I pointed out to Justin that AI is unlike any previous incarnation of edtech and is transforming society in ways unseen in human history. Yet, this is precisely the core of Stephen’s concern. The primary risk is not a lack of technological power, but rather the immense resilience of legacy educational structures. If schools attempt to manage this profound shift purely through the lens of compliance and restrictive policies, they will inadvertently domesticate the technology, reducing a historic societal advancement into just another contested on the edtech shelf.

Justin’s historical warning is precisely why Stephen’s current call to action is so urgent. Stephen recognizes that if schools treat AI as just another shiny tool to stick on the edtech shelf, it will fade into the background noise of failed classroom disruptions. To break the historical cycle and match the gravity of this societal shift, we cannot just add AI to our current classrooms. We have to change our entry point entirely.

Backward Design Matters More Than Ever

Stephen’s answer is not panic. It is purpose.

He returns to backward design: start with the goal. What do we want students to understand? What skills do we want them to build? What kinds of intellectual habits do we want them to practice?

Decide where AI supports that purpose and where it undermines it.

If the goal is simply to produce the “best” paper, AI will increasingly compete with the student. But if the goal is to help students develop the skills needed to produce insightful writing, then the teacher must design for skills such as questioning, source evaluation, argumentation, organization, revision, voice, and judgment.

This is where the teacher’s role becomes more important, not less.

The teacher has to decide what kind of friction belongs in the task. Some friction is unnecessary. AI can help students overcome blank-page anxiety, generate practice questions, receive immediate feedback, or see a concept from another angle. But some friction is necessary and formative. Students still need to wrestle with evidence, confusion, ambiguity, revision, and explanation.

The goal is not less effort. The goal is better effort.

The Human Teacher Becomes More Valuable

One of the most reassuring parts of Stephen’s perspective is that he does not see AI eliminating teachers. If anything, he sees the human teacher becoming more valuable.

But not the teacher who simply ignores AI.

The teacher who matters in this environment is the one with enough disciplinary judgment, pedagogical skill, and AI awareness to design meaningful learning when machines can generate fluent answers on demand. That teacher knows when AI can extend learning and when it can hollow it out. That teacher can look beyond the polished product and ask, "Where is the student’s thinking?” Where is the judgment? Where is the struggle? Where is the growth?

In the humanities, this is especially important because writing has long been both process and product. Students write to show what they know, but they also write in order to discover what they think. When AI can generate fluent prose, teachers need more windows into the thinking behind the prose: annotated sources, oral defenses, revision histories, in-class writing, reflective explanations, and conversations about choices.

Stephen even described feeling reassured by student research papers that were flawed in recognizably human ways. (Remember those days? ) It is a funny and revealing inversion. In an AI-saturated environment, imperfection can become reassuring evidence that the student is still present.

What Schools Should Do Now

Stephen’s work points toward several practical moves:

First, schools should audit major assignments for AI workflow vulnerability. Could AI complete most of this task? If so, what parts of the process need to become visible?

Second, schools should move from tool-based policies to principle-based guidance. A policy that names ChatGPT but ignores search, YouTube, Docs, databases, and embedded AI systems has already aged out.

Third, teachers should define AI roles by purpose. Is AI acting as a tutor, critic, quiz generator, source-finder, editor, simulator, translator, or forbidden co-author? “AI allowed” and “AI banned” are too blunt for the work ahead.

Fourth, schools should make thinking visible before the final product. That may mean process notes, source checks, oral explanations, draft conferences, prompt reflections, comparison tasks, or in-class performance.

Finally, professional development should focus less on tool tours and more on instructional design. Teachers do not only need to know what AI can do. They need time and support to redesign tasks around learning goals, cognitive friction, visible thinking, and authentic evidence of student growth.

The gap between what AI can do and how schools are responding will not close by itself. Vendors will keep building. Students will keep experimenting. Teachers will keep improvising.

The choice for schools is whether to remain reactive—tightening rules around yesterday’s tools—or to do the harder and more hopeful work Stephen Fitzpatrick keeps pointing us toward.

Clarify the purpose.

Redesign the task.

Preserve the thinking.

And keep the human teacher close enough to the learning to know what is really happening.

—

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

They worry that students will use ChatGPT, Photomath, Gemini, or other AI to skip the hard cognitive aspects of mathematics, such as setting up the problem, choosing a strategy, working through uncertainty, checking their reasoning, and learning from mistakes.

But when I spoke recently with Bret Xu, an 11th-grade math teacher in New York City, I heard a different kind of AI story. Bret is not ignoring the risks. He is not pretending AI magically improves math learning. And he is certainly not handing his classroom over to machines.

Instead, he is answering a more interesting question:

What happens when students use AI not to avoid mathematical thinking, but to make their thinking more visible?

Subscribe now

The Problem Was Never Just AI in Math

Bret teaches Algebra 2 at an international high school where many students are newly arrived immigrants from Latin America. As a result, most of his students are Spanish-speaking English language learners, and their mathematical backgrounds vary considerably. Some students are ready for grade-level Algebra 2 work. Others are missing foundational skills from years earlier. Bret described the gap not as one or two years, but in some cases as “eight, ten years apart.” (!)

That creates a teaching challenge both mathematical and linguistic.

Bret does not speak Spanish. Yet, his students are still developing their English abilities. And math, of course, is much more than getting the right answer. It is about explaining and justifying steps, describing patterns, and making reasoning visible. Not an easy task for someone who has been learning English for six months.

So Bret began looking for tools that could help him understand what students were thinking, support students at different levels, and give them more opportunities to explain their reasoning.

Watch my interview with Bret.

Making Math Thinking Visible

One of those tools is Snorkl.

As I’ve written, Snorkl gives students a whiteboard space where they can write, draw, annotate, upload images of handwritten work, and record their voices. The AI then provides feedback based on the task and criteria the teacher has set.

For Bret, the value is not simply that students get feedback faster. It is that they have to explain what they are doing.

In many math classrooms, the final answer receives most of the attention. Students show work, perhaps, but teachers often have limited access to the reasoning behind that work. Did the student understand the concept? Use a memorized procedure? Guess? Copy? Maybe they skipped a step because they did not know how to explain it?

When students speak through a problem, more of that thinking becomes visible.

I was especially struck by an unexpected benefit:

Students were not only working on math, they were developing their verbal English skills.

Bret noted that verbal responses are different from written responses. Once students record their explanation, they cannot simply insert a sentence or quietly revise a phrase. If the explanation is incomplete, they often have to record again. That repetition pushes them to organize their thinking more clearly.

So, students started asking each other “how do I say this better in English”

In other words, the math task becomes a language task, too.

For English language learners, that is powerful. Students are practicing how to explain a process, sequence their reasoning, and communicate logic step by step. They are not just solving for x. They are learning how to say what they did and why it makes sense.

Of course, students do not always embbrace this challenge.

Bret told me that when he announces “Snorkl time,” students sometimes groan. But he does not see the groan as failure. He sees it as evidence that students know they are about to do hard work. They are going to have to think, speak, revise, and try again.

In other words, they participating in cognitive struggle, not cognitive offloading.

Often the assumption is that if students are using AI, they must be avoiding struggle. But Bret’s use of Snorkl suggests the opposite. AI is designed into tasks that increase cognitive demand. It can ask students to explain more, not less. It can make incomplete reasoning harder to hide.

When Scaffolding Becomes a Crutch

But Bret’s experience with another AI tool reveals something different.

Using Playlab, Bret spent a summer building a student-facing math tutor around his own curriculum. He uploaded lesson materials, short videos, scripts, practice problems, homework, and answer keys. Each problem was coded so students working on paper could ask the chatbot for help with a specific problem. The tutor was designed to scaffold, not reveal answers. Students could ask for help. They could check whether their answer was correct. But the chatbot was not supposed to simply give them the final answer.

At first, this sounds close to the dream many teachers have had, a kind of individualized support system that can help every student without making them wait for the teacher to circulate around the room.

But then Bret noticed something.

Even when the chatbot refused to give away the final answer, students began using it too quickly. They would see a problem and immediately ask, “What’s the first step?” Then they would complete that step and ask for the next one. Then the next one.

Technically, they were not cheating.

But they were not struggling productively either.

Hearing this was disappointing but not schocking. It points to one of the central tensions in AI-supported learning: scaffolding can become a crutch.

The danger is not only that AI gives students answers. The danger is that AI gives students the beginning of their thinking before they have tried to begin themselves.

The first step is often a p learning lives.

This is why AI tutor design matters so much. It is not enough to tell a chatbot, “Don’t give the answer.” A strong AI learning system needs architecture. It needs rules for when to help, how much to help, what kind of hint to give, when to ask a question instead of answering one, when to withhold support, and when to tell a student, in effect, “Try for two more minutes.”

That is where machines still struggle.

The Teacher Is Still the Architect

A human tutor can watch a student hesitate and make a judgment. Is the student confused, anxious, lazy, stuck, or almost there? Should I intervene? Wait? Encourage? Redirect? Ask the student to explain what they already know?

A passive chatbot does not naturally make those judgments. It responds. And because it is designed to be helpful, it may become too helpful.

Bret sees this clearly. His next area of exploration is agentic AI — systems that might identify student weaknesses, pull up targeted review materials, assign follow-up practice, and reintroduce concepts through spaced retrieval. If a student struggles with factoring quadratics in Algebra 2, for example, Bret imagines an AI system that can recognize the Algebra 1 gap, direct the student to a short review video, provide practice, and check for understanding before moving on.

That is a different vision from the generic homework-help chatbot.

It is not AI as answer machine. It is AI as part of a learning architecture.

Building Into Interaction

Bret’s classroom is a live experiment.

He is showing what responsible student-facing AI might look like in math class: whiteboard explanations, voice feedback, multilingual support, curriculum-aligned tutoring, teacher-defined criteria, and careful attention to productive struggle.

He is also showing why the design challenge is hard.

AI can help students explain more clearly. It can give teachers better windows into student thinking. It can support multilingual learners in ways that were difficult before. But it can also become a crutch if students learn to ask before they think.

It brings up a pertinent question: What kind of thinking does the AI interaction produce?

Bret Xu is not running from that question. He is building his way into it.

And for any math teacher wondering whether student AI use must inevitably weaken learning, Bret’s classroom offers nuanced optimism: AI can support deeper understanding, but teachers must design the struggle.

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

We hear that AI will transform learning, disrupt assessment, personalize instruction, restructure teacher workloads, threaten academic integrity, prepare students for the future, or change the role of the teacher.

All of that may be true. But too often, the conversation never lands in a classroom: What exactly did the teacher ask students to do? At which point did AI enter (and exit) the task? What did students have to think through? What did the teacher evaluate? What evidence of learning became visible?

Subscribe now

This is a glaring classroom-context gap in much of today’s AI-in-education commentary. The discussions here on Subtack, on LinkedIn, and on other platforms are often thoughtful, eloquent, and well-intentioned. But many feel strangely detached from the actual instructional situations teachers face daily. (Even my posts at times, I must admit.)

“Practical AI” often feels disembodied. It appears as a set of prompts or a list of tools, but not as a classroom scenario. It can be bereft of a student task, a teacher decision, and evidence of what students understood.

To be clear, I don’t mean to trivialize these uses. Teachers are overextended. If AI can help a teacher adapt a reading passage, create practice questions, translate key vocabulary, or scaffold a student who is stuck, I am all for it.

But let’s be clear about what kind of AI use this is.

It’s teacher-facing. It’s about AI-supported teaching.

That’s not the same as AI-supported learning.

Sure, students may benefit when a teacher can use AI to create a lesson plan, simplify a text, generate a revision podcast, build a quiz, or draft a slide deck. They may receive clearer materials, more accessible texts, faster feedback, or better scaffolds.

A Step Toward Concrete Classroom Guidance

A step towards practical, concrete guidance is the UK Department for Education’s recently updated AI support materials.

Unlike much AI commentary, the UK’s Module 4 on generative AI use cases actually lands in teachers' daily classrooms. It is a national, official, public resource designed to help schools and colleges use AI safely and effectively, and it includes slides, videos, transcripts, summaries, activities, and templates. The collection was updated only a week ago, and Module 4 focuses specifically on use cases of generative AI in education.

There is no obvious U.S. equivalent. Since President Trump gutted the Department of Education, American educators have no national, government-backed, classroom-facing professional learning package that offers practical AI use cases for schools and colleges.

So the UK materials are worth examining, not because I think American teachers should adopt them wholesale, but because they show what practical government guidance can look like.

The UK Module 4 Workbook Gets Concrete

Module 4 is of special interest. It is organized around four broad areas: teaching and learning, inclusion and accessibility, administrative tasks, and business and operations. It offers use cases for teachers, support staff, leaders, and operations professionals, rather than limiting AI to a single classroom tool or a narrow productivity frame.

The teaching and learning examples include lesson content and structure, long-term planning, formative assessment, quiz generation, error feedback, adapting text for style or reading age, generating images, creating exam-style questions, producing visual resources, and creating audio resources such as podcasts or verbal explanations.

This is the kind of guidance many teachers are asking for. Not vague statements about AI disrupting teaching, but rather something closer to: Here is how AI might help you prepare tomorrow’s lesson, adapt this text, support this learner, generate practice questions, or make a resource more accessible.

Inclusion and Accessibility is the strongest section. The workbook lists AI uses such as breaking down tasks, summarizing, explaining concepts further, creating exemplars, building writing frames and sentence starters, translating content, differentiating by reading age, using visual and audio materials for dual coding, converting audio to text, and creating audio descriptions of images.

In all, it’s grounded in the real work of helping students access material.

The UK guidance also repeatedly asks educators to consider pedagogical need, contextualize the example for their setting, align use with teaching and learning policy, protect student data, check outputs for accuracy and bias, and adapt AI-generated materials before using them with students.

That combination — concrete examples plus professional judgment — is exactly what many schools need.

But these same examples also reveal an important limitation.

Practical Usually Means Teacher-Centric

Most of the UK guide’s practical examples follow a familiar pattern:

The teacher prompts AI.
AI creates or adapts something.
The teacher checks the output.
Students use the resulting material.

As I’ve said, the result is often valuable. A simplified text can help a student participate. A translated vocabulary sheet can support access. A scaffolded essay plan can reduce cognitive overload. And a visual description can open up material that might otherwise be inaccessible.

But students are typically the recipients of AI-supported materials rather than active participants in AI-supported learning.

Practical usually means: How can AI help teachers do their work?

This starting point is completely understandable. Teachers are focused on planning, instruction, differentiation, feedback, communication, and the many small decisions that keeps their classrooms going. If AI is going to enter schools responsibly, teachers need low-risk ways to understand what these systems can and cannot do. Teacher-facing use also builds familiarity. It helps educators experience the strengths, quirks, errors, and limitations of AI before asking students to use it.

But if practical AI stops there, AI mostly remains in the preparation layer.

It helps teachers prepare materials, adapt resources, and manage workload. It may improve access. It may make some instruction more responsive. But it does not necessarily change what students do as learners.

Washington Adds the Student-Use Layer

Washington State’s Human-Centered AI Guidance for K–12 Public Schools helps articulate what the UK guide does not fully develop.

Washington’s guidance is not as rich as the UK materials in classroom-ready examples. A teacher looking for dozens of immediate use cases will find the UK Module 4 more useful.

But Washington asks a different question:

How should schools structure AI use so students remain active, ethical, reflective learners?

Its central frame is “Human Inquiry → AI → Human Empowerment.” AI use should begin with human inquiry and end with human reflection, human edits, human understanding, and human empowerment. The guide explicitly warns against letting AI become the final thought, product, or paper.

That’s an excellent instructional frame. As a result:

The question goes beyond simply whether students used AI for learning. The question is what role AI played in the student’s thinking.

Of special note is the section titled “Implementing AI: A Practical Guide for the Classroom.” Washington’s five-step AI scaffolding scale helps teachers move beyond the weak binary of “AI allowed” or “AI banned.” The scale runs from no AI assistance to AI-assisted brainstorming, AI-supported drafting, AI-collaborative creation, and AI as co-creator. At each level, the guidance emphasizes student contribution, revision, transparency, citation, critical evaluation, rationale, and academic integrity.

This is a point where many schools are struggling. AI might be appropriate for brainstorming but not drafting. It might be useful for feedback, but not drafting. It might generate counterarguments, but the student still has to evaluate which ones are weak, relevant, or worth addressing. It might suggest multiple solution paths, but the student still has to explain which path makes sense and why.

While I would suggest a slow buildup to Level 4 or 5, Washington’s assignment scaffolding matrix extends this idea across reading and discussion preparation, long-form writing, creative writing, podcasts and videos, research, and problem solving.

That shifts AI guidance from policy into assignment design.

A school policy can say, “Students must cite AI.” A teacher still has to decide what AI use means in a history essay, a biology research project, a Spanish dialogue, a math problem set, or a student-produced podcast.

What kind of AI help is appropriate? What kind would undermine the learning? What process evidence should students produce? How will the teacher know what the student understands?

These are the practical questions schools cannot avoid.

From Detection to Visible Process

If the final product is the only evidence of learning, AI will always create uncertainty. If the process is visible, AI becomes easier to discuss, evaluate, and teach with.

Visible process might include prompts, drafts, annotations, revision notes, reflections, oral explanations, teacher conferences, version history, or disclosure statements. Students might be asked to explain which AI suggestions they accepted, which they rejected, and how their own thinking changed.

In all, this is a step towards better learning design:

• A history assignment in which students use AI to generate three possible interpretations of a primary source, then identify omissions, unsupported claims, and missing context before writing their own interpretation.

• Or an English assignment in which students use AI to generate counterarguments to a thesis statement, then annotate which counterarguments are valid, weak, or irrelevant before revising their essay.

• Or a math task in which students ask AI for two different solution paths to the same challenging problem, then compare the methods, explain which one is more efficient or easier to understand, and then solve a similar problem without AI support.

• Or a science investigation in which students ask AI to critique an experimental design, then revise their method and explain which suggestions they accepted or rejected.

• Or a world language role-play in which students use AI voice practice, save corrections, re-record, and reflect on grammar and pronunciation. (Something which I do as a language learner.)

These are the kinds of classroom cases we need more often.

A student who can prompt AI but cannot critique AI is not AI literate. Practical guidance should help students check accuracy, identify bias, notice missing perspectives, recognize hallucinations, question overconfident claims, and decide when AI output is not good enough.

Cuban’s Warning Still Looms Large

More than a year ago, I wrote about Larry Cuban’s warning that AI might simply be adapted to the existing contours of schooling — used at the margins rather than forcing educators to rethink how students learn. That warning still applies. If practical AI guidance is limited to teacher productivity — lesson plans, quizzes, emails, summaries, slide decks — then AI may become common in schools while remaining peripheral to learning.

Combined, the UK DfE materials and Washington State guidance suggest a better path. One brings AI down into the daily work of schools. The other helps schools structure student AI use so that thinking, judgment, and academic integrity remain visible.

Practical AI guidance should include both.

It should show what teachers can do with AI: plan, scaffold, adapt, translate, give feedback, communicate, and improve access. But it should also show what students do with AI: question, critique, revise, disclose, explain, and take responsibility for their work.

A practical AI guide should not stop at “Here is what the teacher can generate.” It should show what students do, what AI does, what the teacher sees, and where the learning happens.

The most practical AI guidance will help teachers decide not only what AI can produce, but what students must still think through, struggle with, explain, revise, and own.

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

Google I/O 2026 introduced or previewed a range of AI-powered tools and capabilities, which will undoubtedly impact schools enmeshed in the Google ecosystem. They include: Gemini Spark, an agentic assistant designed to complete multi-step tasks; Docs Live, which turns spoken ideas into structured documents; Ask YouTube, which makes video search more interactive, and Gmail Live, which lets users converse with their inbox. Also introduced were new AI-powered design and image tools; more advanced search experiences; and learning-oriented features such as guided learning, study guides, and test prep.

https://blog.google/innovation-and-ai/technology/ai/google-io-2026-all-our-announcements/

Tom’s Take:

New $11M Effort Aims to Train Teachers in AI. How Does It Work?

Education Week reports that the Computer Science Teachers Association is launching a two-year, $11 million multistate initiative to prepare thousands of K–12 teachers to teach foundational computer science and AI. The program, called AI PD Weeks, begins this summer in Illinois, Indiana, Iowa, Minnesota, New Jersey, and South Carolina, with four more states planned in year two. It will offer weeklong summer training, stipends for continued school-year PD, and research on how teachers integrate AI concepts, tools, and ethics into instruction.

https://www.edweek.org/technology/new-11m-effort-aims-to-train-teachers-in-ai-how-does-it-work/2026/05

Tom’s Take: I applaud efforts to move AI PD beyond scattered webinars and vendor tutorials into a structured, research-backed professional learning effort. The devil is always in the details, but this initiative should fortify the link between AI literacy and foundational computer science education.

UK Department for Education AI in education support materials

The UK Department for Education’s “Using AI in education settings: support materials” collection includes four guidance modules for all school and college staff—covering understanding AI, interacting with generative AI, safe use, and use cases—plus separate leader-focused support for school and college leaders. The materials include slides, videos, summaries, activities, and templates, and can be adapted for individual or group use at different experience levels.

https://www.gov.uk/government/collections/using-ai-in-education-settings-support-materials

Tom’s Takes: This is significant because it’s official national guidance, not commentary or vendor PD. It’s also a signal that effective use of AI is a governance and professional learning priority. I jumped straight to Module 4: Use Cases and found some excellent, practical examples.

“More Schools Are Providing AI Training for Teachers. Is It Any Good?”

Education Week reports that K–12 teacher AI training has expanded quickly, but unevenly. The article says the share of teachers reporting no AI training fell from almost 60% in October 2024 to 50% by fall 2025, while a clear majority now report receiving at least some training. At the same time, the piece says the frequency, depth, and quality of AI PD vary widely, and many teachers remain uncertain or unenthusiastic about learning to use AI. The article questions how meaningful the training really is.

https://www.edweek.org/technology/more-schools-are-providing-ai-training-for-teachers-is-it-any-good/2026/05

Tom’s Take: Only 50? There’s an urgent need for AI PD, especially now that we’ve entered the age of agentic AI. But most important is whether that training actually changes teaching and learning. This article creates the impression that professional development is scattered and shallow — and that’s worrisome.

“From AI Tutors to AI Study Mates”

Dr. Philippa Hardman argues that many AI tutors are built around task completion rather than durable learning, creating what she calls the learning-performance paradox. The piece suggests moving away from answer-giving AI tutors toward “AI study mates” that support productive struggle, metacognition, and contextual learning. Hardman says most current tools still fail to build a real evolving picture of the learner over time, and profiles examples such as RiPPLE, Recast, and Khanmigo to show what more pedagogically grounded AI companions can and cannot yet do.

Dr Phil's Newsletter

From AI Tutors to AI Study Mates

Hey folks 👋…

Read more

3 months ago · 59 likes · Dr Philippa Hardman

Tom’s Take: As usual, Dr. Hardman’s analysis, argumentation, and presentation are excellent. But her framing strikes me as a little too narrow. The more important shift is not simply from answer-giving tutor to study companion, but toward interaction models that can identify hesitation, self-correction, partial understanding, and make student thinking visible to teachers.

America’s Schools Are Terrible at Catching Kids Up. How AI Can Help

Daniel Weisberg argues that one of American education’s deepest problems is not simply low performance, but schools’ failure to help struggling students catch up once they fall behind. Drawing on TNTP research and concerns about incoming student preparedness at UC San Diego, he says teachers are being asked to do the impossible without timely diagnostic insight or strong intervention tools. He presents AI as a possible support system that can help surface real-time learning patterns, identify misconceptions, and suggest evidence-informed next steps for teachers.

https://www.the74million.org/article/americas-schools-are-terrible-at-catching-kids-up-how-ai-can-help/

Tom’s Take: Weisberg offers a systems-level approach to tackle one of education’s most challenging problems. Instead of positioning AI as a productivity tool, Weisberg portrays AI as a means to strengthen diagnosis, intervention, and teacher decision-making.

And More:

• Glaze: New Vibe Coding for Beginners - Jeremy Caplan

• Free Graded Pro in exchange for taking part in a research project

• What is Edcafe AI and How Can I Use It To Teach - Tech & Learning

• Elevating the Classroom: A Blueprint for Strategic AI Integration and Student Empowerment - Tech&Learning

• Students Are Digital Natives. Let Them Lead the AI Revolution in Education - The 74

• When Highlights Are Easy to Fake With AI, Integrity Matters More - Tech&Learning

• Philadelphia Middle Schoolers Explore How AI Changes Their Classrooms and Their Lives - The 74

• Google expands its educator-and-learner AI initiatives in India. - Google

• How to Start a STEM Club for Girls - TCEA

• The Two Faces of AI in Education - Richard Noble

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

Google introduced or previewed a range of AI-powered tools and capabilities: Gemini Spark (see clip below), an agentic assistant designed to complete multi-step tasks; Daily Brief, which synthesizes email, calendar, and tasks; Docs Live, which turns spoken ideas into structured documents; Gmail Live, which lets users converse with their inbox; Ask YouTube, which makes video search more interactive; new AI-powered design and image tools; more advanced search experiences; and learning-oriented features such as guided learning, study guides, and test prep.

Taken individually, some of these developments seem a tad removed from the classroom. Some may not reach school accounts for months — if at all. Some may remain tied to paid consumer subscriptions, age restrictions, admin controls, or U.S.-only rollouts.

But taken together, they point to something momentous:

Google I/O 2026 shows that AI is becoming the operating layer of school work. The question for educators is whether we will redesign teaching, assessment, and student support enough to keep learning — not automation — at the center.

That’s the challenge.

Subscribe now

Keep in mind that Google does not need to “enter education” in the way a new edtech startup does. Teachers already write in Google Docs, store materials in Drive, communicate through Gmail, present with Slides, assign through Classroom, search with Google, show videos through YouTube, and teach on Chromebooks.

So when Google changes the AI capabilities around Search, Docs, Gmail, YouTube, Gemini, Drive, and eventually Google Classroom-adjacent workflows, it is not just adding another element to a school technology repertoire.

It is changing the school’s work environment.

From chatbot to agent

Dan Fitzpatrick captured the central shift in his Forbes piece on Google I/O 2026: “Gemini is no longer a chatbot you ask things. It’s an agent that goes and does things.” He’s talking about Gemini Spark, Google’s new agentic assistant. Your personal AI assistant.

It’s no chatbot. It’s a different category of AI use:

Spark will be able to coordinate entire student workflows. It can search, retrieve, summarize, draft, compare, organize, and take action across tools. For a school assignment, that means a teacher’s question shifts from “Did AI write the student’s paper?” to “Did AI complete the assignment’s entire workflow?”

If Spark — or something similar — can perform the major steps of a task, then we need to ask what the task is actually measuring.

If students can use AI to gather resources, summarize information, generate slides, create visuals, draft scripts, and prepare polished papers, then educators must make student thinking more visible. We need to see how students framed the problem, what sources they trusted, what they changed, what they discarded, what they really understood, and what they can explain without AI doing it for them.

I am not saying we should panic. I’m saying we need better design and assessment.

Yes, the finished product still matters. But it can no longer be accepted as the only evidence of learning.

The teacher as manager of AI rhythms

One of the more interesting implications of Google’s direction is that the teacher’s role may become less about producing every instructional artifact and more about orchestrating teaching rythms.

Let me illustrate this idea: A teacher might use Gemini Spark to create three versions of a reading passage, generate discussion questions, produce a visual explanation, organize student misconceptions, draft a formative assessment, and prepare a quick reflection prompt for the next class.

Individually, teachers could do all of those things before.

But now, using an AI agent, the teachers does them quickly, repeatedly, and in response to immediate student needs. This changes the instructional rhythm. The teacher becomes an on-the-spot designer, editor, and orchestrator. In theory, in enables teaching to become more adaptive, more flexible and more iterative.

But it really comes down to how we define “teacher-led.” In Fitzpatrick’s interview with Google’s Chris Phillips, VP and General Manager of Geo and Education, Phillips repeatedly emphasized that Google’s learning-specific demos were “still led by the teacher.”

Educators should welcome that language — and hold Google to it.

Google is trying to frame this AI layer as teacher-led, confidence-building, and learning-centered, but schools need to define what that means before vendor language becomes school language.

To be clear, “teacher-led” must never degenerate into teacher clicking buttons before AI generates everything. It must translate into teachers adapting, constraining, questioning, revising, and redirecting the AI. Teachers need to remain close enough to the learning process to understand what students are doing and why.

But will that happen?

The problem is not that AI helps teachers multitask and work faster. The problem is that the AI may quietly usurp instructional decisions.

If AI does the heavy lifting of intellectual preparation, does the teacher somehow lose their connection to the material? Instructional planning decisions come from wrestling with the content while building the lesson. If we automate the friction of preparation, do we inadvertently automate away the teacher's expertise?

The quiet procurement story

In his Forbes article, Fitzpatrick highlights a point that may matter enormously for school systems: cost. He argues that Gemini 3.5 Flash may not sit at the frontier of model benchmarks, but if it is fast, affordable, and “good enough” for most classroom tasks, then it changes the economics of AI adoption.

That may be.

Chromebooks became a popular low-cost 1:1 alternative to laptops because they were “good enough” to support ordinary educational work at scale.

But Chromebooks have fundamentally changed what schoolwork looks like, moving everything to the cloud, and (with Google Classroom) making the school work environment highly standardized.

A similar process may unfurl with with AI. Schools do not need the most powerful AI model in the world for every use case. They need reliable, affordable, safe, manageable AI that can support regular educational work at scale.

A fast, lower-cost model can help teachers draft materials, summarize information, generate practice questions, support feedback, and create learning supports. If those capabilities become inexpensive enough, AI adoption may accelerate not because schools are chasing some tech ideal, but because useful AI becomes embedded in tools they already subscribe to.

So, how will this standardize the school work environment?

Oh, and affordability is only one side of the access question.

Consider: Which features will be available in Google Workspace for Education? Which will require consumer subscriptions? Which will be available to students under 18? What controls will administrators have? What data protections apply? Can a school turn features on or off by grade level? Will wealthy students with personal AI subscriptions have access to more powerful agents than classmates using school-managed accounts?

The future may be cheaper and yet persistently unequal.

Search, YouTube, and the new learning interface

Some of Google’s most important education implications may come from tools not branded as education tools.

If Search can generate interactive explanations, simulations, timelines, and visualizations around a student’s question, then search is no longer just a list of links. Search becomes a custom learning interface.

If Ask YouTube helps students ask a question and jump directly to the relevant part of a video, then YouTube becomes more tutor-like. Students will not just browse videos; they will interrogate them.

If Docs Live lets a teacher speak rough ideas into a phone and have Gemini structure them into a usable document, then lesson planning becomes more conversational.

But the same capability raises hard questions for assessing student work.

If a student can ramble a few half-formed thoughts into a phone and have AI turn them into a structured essay, lab reflection, or project proposal, what exactly are we assessing? The student’s understanding? Their ability to organize ideas? Their oral communication skills?

On the bright side, students may receive more accessible explanations, more visual supports, and more personalized pathways into complex topics. But they may also become more dependent on AI-curated summaries, pre-selected clips, and generated representations. Reading between the lines, comparing sources, wrestling with ambiguity, and building one’s own understanding still matter tremendously.

What educators should do now

Regardless of whether these products reach schools in weeks or months, educators can’t ignore the huge wave coming:

School leaders should audit major assignments. Where could an AI agent complete the workflow? Where do students need to show process, reasoning, source judgment, revision, or oral explanation?

Teachers should design for visible thinking. Ask students to document prompts, explain revisions, identify what they verified independently, and reflect on how AI changed their work.

Schools should review their Google ecosystem. If your school runs on Google Workspace, then your AI policies needs to include Workspace, Search, YouTube, Drive, Classroom, and Chrome — not just ChatGPT.

Most importantly, school leaders should treat AI literacy as part of everyday learning, not a separate assembly or one-time policy document. Students need to understand how AI generates, distorts, omits, and persuades.

From destination to layer

Phillips points to Google’s Educator Series as one vehicle to help teachers build confidence in small modules. But schools should not outsource their AI learning strategy to Google. They need in-school conversations about assessment, student use, privacy, equity, and what responsible AI use looks like in their own classrooms.

Google I/O 2026 provides a push.

AI is clearly moving from a separate destination to an operating layer. It is entering the tools teachers and students already use. It is becoming more agentic, multimodal, affordable, conversational, and embedded.

The question is whether we can use these tools to make learning more visible, more thoughtful, more personal, and more human.

That considerable challenge belongs to educators.

—

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

The Google for Education Impact Team has released the first content of its new Google AI Educator Series. The 60-minute video session explains the 31 new bite-sized, stackable learning sessions designed to transform how educators interact with AI using Gemini’s free tools and capabilities.

Tom’s Take: “Part 2: K-12 Spotlight” is the most useful section for K12 educators. It explores explore the 15-minute session model, covering grade-band specific pedagogical tools from early literacy to post-secondary pathways.

EL teachers’ experiences show how messy classroom AI use really is

Larry Ferlazzo’s opinion essay “What in the ChatGPT Is This?: How EL Teachers Are Navigating AI Use” provides three teacher-grounded accounts of how English learner educators are actually working through AI in practice rather than in abstraction.

https://www.edweek.org/technology/opinion-what-in-the-chatgpt-is-this-how-el-teachers-are-navigating-ai-use/2026/05

Tom’s Take: ELL teachers often face some of the most vexing challenges when it comes to scaffolding, language support, equity, and overreliance. These interesting teacher commentaries point to the real pedagogical tradeoffs they face.

EduGems: Study Coach Gem Creator

Eric Curts’ Study Coach Gem is a customizable AI tutor built in Google Gemini that helps students review content, practice retrieval, and build AI literacy through guided conversation. Its companion Creator tool asks teachers five questions about their classroom context, then generates ready-to-use instructions for a reusable Gem in about three minutes.

https://www.edugems.ai/gem/study-coach-gem-creator

Tom’s Take: Eric Curts excellent EduGem frees teachers from writing complex instructions, or paying for custom design, to create a free, customizable and reusable classroom-specific study assistant through Google Workspace for Education.

Have We Reached a Tipping Point on Screens in Schools?

Stephen Fitzpatrick argues that the screen backlash in schools is real and often justified, especially in K–8, but warns against treating device removal as a cure-all. He says many districts rushed into 1:1 Chromebook adoption during the pandemic without developing strong pedagogy, guardrails, or infrastructure for responsible use. Rather than blanket bans, he advocates more targeted rollbacks, including sharply limiting screens for younger students, tightening high-distraction platforms like YouTube, preserving device access for older students, and building more thoughtful, developmentally staged approaches to digital and AI literacy.

Teaching in the Age of AI

Have We Reached a Tipping Point on Screens in Schools?

In March of 2020, millions of teachers and students across the country pivoted online for distance learning because of the pandemic. Districts that did not have 1:1 laptop programs swiftly invested in Chromebooks for every child. Despite the imperfect nature of the experience, we would not have been able to teach or connect with our students without tho…

Read more

4 months ago · 25 likes · 20 comments · Stephen Fitzpatrick

Tom’s Take: Fitzpatrick provides a thoughtful approach to selective roolbacks of screen use. But following his recommendations is another matter; schools continue to struggle to build thoughtful, developmentally staged approaches to digital and AI literacy.

“11 AI Prompts Every Teacher Should Know”

This guide from Jeremy Caplan in The 74 offers practical prompt patterns aimed at saving teacher ’ time, sparking engagement, and helping students learn. The article explicitly positions prompting as a workflow skill for classroom professionals rather than as a generic novelty. It reflects a pragmatic, educator-facing strand of AI commentary: less about big theory or policy, more about usable routines teachers can apply to everyday classroom and planning tasks.

https://www.the74million.org/article/11-ai-prompts-every-teacher-should-know/

Tom’s Take: Caplan is the director of teaching and learning at CUNY’s Newmark Graduate School of Journalism and author of the Wonder Tools newsletter. In this piece, Caplan frames prompting as a practical literacy for planning, differentiation, and classroom support and provides a series of useful examples.

Vriti Saraf on Reimagining the Teacher’s Role for the AI Age

In a Getting Smart podcast, Vriti Saraf of Ed3 argues that AI should be used to redesign teaching, not just automate it. Drawing on Ed3’s new research, the conversation says schools risk using AI only for efficiency or test-score gains when the larger opportunity is to free educators for deeper relational, advisory, and design-centered work. Saraf also suggests the future of teaching may be more team-based and role-differentiated, with new educator personas such as learning architect, life navigator, and community connector.

https://www.gettingsmart.com/podcast/vriti-saraf-on-reimagining-the-teachers-role-for-the-ai-age/

Tom’s Take: What caught my attention is the discussion of “future roles” for teachers. Jump to 22:30 for a take on teachers as learning architect, life navigator, and community connector.

AI may make structured literacy more scalable without replacing teachers

This THE Journal article argues that purpose-built AI systems can analyze student-performance patterns, identify specific skill gaps, and connect those gaps to instructional recommendations. The effect, when done well, is not to remove the teacher but to sharpen the teacher’s ability to act. The piece positions AI as an instructional support layer within literacy work rather than a substitute for teaching.

https://thejournal.com/articles/2026/05/13/ai-wont-replace-teachers-but-it-may-be-what-makes-structured-literacy-work-at-scale.aspx

Tom’s Take: AI is valuable when it amplifies teacher action in high-need areas like literacy, not when it replaces relationship or instructional judgment. This is among the many instances where AI can amplify the teacher, but that largely gets lost in the prevailing and binary “AI or teacher” debate.

“The 20% AI Policy Won’t Work. Here’s What May.”

TCEA argues that policies such as “your work may contain up to 20% GenAI content” are unenforceable because they assume AI detection can reliably measure such use. The piece says that assumption has been wrong for a long time. It frames the problem as a mismatch between policy language and technical reality, and pushes educators toward clearer assignment-level expectations and more workable integrity practices rather than arbitrary percentage caps.

https://blog.tcea.org/the-20-ai-policy-wont-work-heres-what-may/

Tom’s Take: It’s hard to govern AI when policies can’t be implemented fairly. More emphasis should be placed on assignment design, disclosure, and teacher judgment and less on unreliable detection.

Teacher MOVE

Complete a Google AI Educator Series session, then redesign one real workflow the same day.

Follow the Google AI Educator Series. After the next session, pick one recurring small task you already do in Google Workspace for Education—lesson adaptation, quiz creation, parent communication, review materials, or study support—and redesign that task with Gemini. Start small and build for real. Test whether a Google-supported workflow actually improves your teaching, or gives students better support.

And More…

• Introducing Googlebook, designed for Gemini Intelligence - Google

• Using Gemini AI To Prepare For Standardized Tests - Tech&Learning

• In a new survey, AI scores high as a math learning tool - eSchoolNews

• A Cyberattack on Canvas Could Cause Lasting Aftershocks for Schools - Education Week

• Anthropic, Gates Foundation launch $200 million partnership for AI in health, education - Reuters

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

This strikes me as much more than a user-interface improvement.

For education, it signals a breakthrough in learning with AI that hinges on an enhanced interaction model.

When we imagine a student learning with AI we typically picture the student asking a question, and then the AI answering the question. The interaction still largely revolves around content delivery. The AI may be faster and more personalized than, say, a textbook or video, but the basic structure remains the same. The student seeks information. The machine provides it.

Subscribe now

But what if the AI is not simply waiting for a question?

But what if the AI can recognize hesitation, self-correction, confusion, or partial understanding? What if it can ask a clarifying question before a misconception solidifies? Or, just be quiet because the student is thinking?

That’s a live interaction layer for learning. And it goes well beyond the prevailing “AI tutoring” concept.

Thinking Machines argument is that prevailing AI interactions force humans to adapt to the machine. Users stop, formulate a prompt, wait, receive a response, and then decide what to do next.

But that’s not how many human learning conversations work.

From Content Delivery to Process Observation

Say a student is learning Spanish. The student is telling a story in the past tense and says, “Ayer yo voy… fui… fui al parque con mi amigo.” A traditional chatbot might wait until the student finishes and then provide a grammatically correct version. A better AI tutor would explain the difference between voy and fui.

But a better interaction model notices the self-correction in the moment and says: “Good. You corrected yourself from present to preterite. Now try the next example and pay attention to the same pattern.”

That might seem like small difference, but it is pedagogically important. The AI is not simply correcting user output. It is addressing a learning process.

• In math, a student might talk through a solution while writing steps on a screen. The AI could notice that the spoken reasoning is sound but the written notation is inconsistent. It then focuses on this inconsistency.

• In science, a student might describe a lab result while pointing to the wrong part of a graph. The AI points this out and asks why.

• In history, a student might be describing the importance of a cultural site. The AI immediately produces visual evidence that supports, or challenges, the student’s assertions.

• In writing, a student might verbally explain an argument, while the AI targets claims and gaps.

In each case, the educational importance is not just that AI can provide an answer. The value is that AI can help make student thinking visible.

Summative assessments by themselves often lead teachers astry. A finished essay can hide the process that produced it. A completed math answer can hide the misconception. A polished AI-assisted response can hide whether the student understood anything at all.

Interaction Models could make the learning process more observable. They could explore a hesitation, a revision, a spoken explanation, a visual reference, and a moment of self-correction. That is far more instructionally useful than another polished answer.

The Giants are Moving: Google’s Multimodal Bet

Google may not be using the same terminology as Thinking Machines, and I wouldn’t claim that it is building the same kind of model. But Google’s recent education work seems to point in a similar direction. LearnLM, Guided Learning, Gemini Live, Project Astra, NotebookLM, and Google Classroom integrations all suggest that Google’s tutoring bet is not simply about creating a chatbot that explains things better. It is about embedding AI more deeply into the flow of learning.

Guided Learning already nudges Gemini away from quick answers and toward questioning, step-by-step support, diagrams, quizzes, and multimodal explanation. Project Astra points toward live AI that can see what the learner sees, respond to the surrounding context, and interact through voice and visual input. NotebookLM is becoming more of a representation layer, able to turn source material into slides, reports, study guides, audio, video, and visual outputs. Google Classroom integrations suggest that these capabilities may increasingly sit inside the environment where teachers already assign, organize, and assess student work.

And Google has “realized that as new capabilities in AI are emerging and also maturing — for example — we have these demos of live experience where it’s kind of video and audio and you essentially can just talk to AI in the same way as you would talk to your human teacher,” says Irina Jurenka, the research lead for AI in education at Google DeepMind.

Put these pieces together and it certainly points to enhanced interaction models with layers distributed across its learning ecosystem.

Amplification, not Replacement

It also changes the role of the teacher. If AI can support more simultaneous practice, teachers may spend less time being the only source of immediate feedback and more time designing the conditions under which feedback happens. Teachers will need to decide: What should the AI notice? When should it interrupt? What counts as productive struggle? Should feedback focus on accuracy, reasoning, confidence, fluency, creativity, or revision? What should be visible to the teacher? What should remain private to the learner?

Imagine a classroom where twenty students are practicing oral explanations using visual aids with an AI partner. The teacher cannot listen to and watch every interaction in real time. But the system might reveal that many students are avoiding evidence, several are misusing a key term, a few are using visual aids misleadingly, and some are simply winging it. The teacher can then respond to the class or redesign the next activity.

That’s not AI replacing the teachers. That’s AI amplifying the teacher.

Yes, the risks are real — and tricky. More human-like AI may also become more persuasive, more emotionally attractive, and harder for students to evaluate. If an AI can interrupt, encourage, redirect, and challenge, then schools need clear boundaries. Who controls the interaction? What data is captured? Are voice, facial, or behavioral signals being analyzed? Can students pause the AI? Can teachers inspect the interaction?

The more natural the AI feels, the more important these questions become.

In all, Interaction Models are not some utopian ideal, nor to be feared. It’s all part of a significant shift. AI in education is moving beyond the chatbot box. The future is more than smarter LLMs. It is better interactions.

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

.

Education Week reports that resistance to school technology has moved beyond scattered complaints and become a broader cultural and political backlash. The article links that shift to three developments: the rapid expansion of one-to-one devices during and after the pandemic, the rise of generative AI, and growing alarm about student mental health and screen exposure. It cites EdWeek Research Center data showing mixed views among educators. Many still see academic value in ed tech, but many also believe it harms behavior, well-being, and social-emotional development.

https://www.edweek.org/technology/the-ed-tech-backlash-is-here-what-it-means-for-schools/2026/04

Tom’s Take: The number of negative articles regarding AI in education has been growing in my inbox for the last several months. An AI backlash is understandable, and not very surprising, but the impulse to ban tech in schools is misguided and worrying. Many criticisms misleadingly link screen time to decreased student achievement and ignore the many educational benefits of tech integration.

Subscribe now

Commentary: Screen Restriction Is Not Pedagogical Reform

Nick Potkalitsky argues that recent debates over screen restrictions and phone bans are being miscast as pedagogical reform. Drawing on early phone-ban research, he says the findings are modest and specific, not evidence that screen caps by themselves improve teaching and learning. His point is that restricting devices may change certain conditions of school life, but it does not automatically produce a better curriculum, stronger instruction, or richer intellectual work. In other words, limiting screens is a governance move, not a substitute for educational design.

Educating AI

Screen Restriction Is Not Pedagogical Reform

Educating AI publishes twice weekly. If this piece was useful, share it with a colleague who’s navigating these questions. And if you want to support the work, consider becoming a paid subscriber. It’s what keeps this going…

Read more

4 months ago · 20 likes · 5 comments · Nick Potkalitsky

Tom’s Take: Potkalitsky correctly points out that device regulation is actually from actual pedagogical reform. Bans and caps may address some behavioural problems, but they do not, by themselves, create better teaching.

NYC Cancels Plans for AI-Focused School, Upper West Side Schools

New York City Schools canceled plans to open Next Generation Technology High School, a proposed AI-focused selective high school in Lower Manhattan, and also withdrew related proposals affecting several Upper West Side schools. GovTech reports the reversal came after months of community opposition and just hours before scheduled protests. Chancellor Kamar Samuels said the proposals were being removed from consideration for now and that the city would consider stronger parent and community engagement before moving forward with future opening, closing, or reorganization plans.

https://www.govtech.com/education/k-12/nyc-cancels-plans-for-ai-focused-school-upper-west-side-schools

Tom’s Take: High-profile AI-themed school innovation in New York, Texas, and other parts of the country is not immune to ordinary school politics. Even a high-profile AI-focused campus initiative can be stopped dead in its tracks over concerns about admissions, transparency and equity. As the AI backlash grows, it may be harder than ever for education leaders to gain public trust in tech-infused initiatives.

An Educator Explores Agentic AI

Lance Eaton explains agentic AI as a newer class of tools that can act within digital systems after a user provides instructions and permissions, going beyond standard generative AI. He describes experimenting with tools such as ChatGPT, Codex, and Claude Code to organize folders, rename and sort research PDFs, build simple apps, process video files, and create personalized learning spaces. At the same time, he stresses caution, backup practices, and a key unresolved question: where agentic AI meaningfully fits teaching and learning.

AI + Education = Simplified

Exploring Agentic AI

I’ve been trying to do this post for a while, and at this point, I think it’s better to have a half-thought post to start the conversation than a fully fleshed one. I want to share a bit about my own usage and way-finding through some of the agentic AI tools I’ve been playing with. I shared a bit about what I’m doing in…

Read more

4 months ago · 36 likes · 4 comments · Lance Eaton

Tom’s Take: Eaton demonstrates how agentic AI is genuinely different from ordinary prompting, especially in terms of productivity and workflow design. At the same time, he raises a key educational concern: whether tools built to remove friction can support learning without undermining it.

Coursemojo Aims to Help All Students Find Their “Reading Mojo”

Coursemojo is an AI startup led by Dacia Toll that aims to support students and teachers in English Language Arts. In this The 74 interview, she frames the company as using artificial intelligence to help students find their “reading mojo” while also supporting teachers. She points to very weak reading results and says too many students are “falling off a cliff” in literacy, and she sees Coursemojo as worthy because it tries to improve the instructional loop, not just add AI on top of classroom work.

https://www.the74million.org/article/the-ai-startup-aiming-to-help-all-students-find-their-reading-mojo/

Making Graphic Novels with NotebookLM

And More…

• Goblins - app for math learning

• Making of the Nation: America at 250 - Google Arts and Culture

• ChatGPT for Excel and Google Sheets - ChatGPT

• Mira AI glasses: A “second brain” - Mira

• Largest phone-ban study finds mixed results for schools - The 74

• Recap of Google’s major April AI announcements - Google

• Claude is now generally available for 365 - Claude

• Microsoft, Google and xAI to give US government early access to AI models for security checks - Reuters

• Big Tech’s AI spending and cloud growth accelerates - Reuters

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

When Google announces a new tool or feature, access typically varies by account type, Workspace edition, admin settings, rollout timing, country, and language availability. Too often, I simply cannot see a feature fully in my own Google account.

So, I take to watching AI video tutorials. A good tutorial lets me see the clicks, the interface, the prompts, the outputs, and the limitations. It helps me move from “Google says this is available” to “Here is what a teacher might actually do with it.”

Subscribe now

Gemini File Generation

One update that caught my attention is Gemini file generation. On April 27, Google announced that Gemini can now create formatted files directly from chat, including Google Docs, Sheets, Slides, PDFs, Word documents, Excel files, CSVs, LaTeX, text files, RTF files, and Markdown files. Google describes the feature as a way to move from “conversation to creation,” producing functional and downloadable files without leaving the Gemini app.

I tried it and had a blast turning research reports into a slideshow and spreadsheets into PDFs.

And it has big implications for a classroom. A prompt might become a lesson plan, quiz draft, rubric, vocabulary list, spreadsheet template, slide deck outline, or student handout. But of course, the important question is whether the file is instructionally useful enough for a teacher to revise and adapt.

Teacher’s Tech video on Gemini’s new file feature is helpful in answering that question. Jamie Keet is not presenting specifically as a classroom teacher, but thoroughly tests the feature in action (unlike a Google promotional video). He shows what Gemini can generate, what the files look like, where the feature works well, and where it has a big limitation. It can help move a teacher from “Gemini can create files” to “Hmm. I could use this to create resources for my students!”

Gemini in Google Classroom

Google has also been expanding Gemini inside Google Classroom, with tools designed to help teachers create content, brainstorm ideas, and differentiate resources.

Any update to Google Classroom is big news because is where many teachers organize assignments, distribute materials, communicate with students, and manage day-to-day instruction. But the question again becomes practical: Does it help teachers with the work they already do, or does it become one more tool to manage?

I don’t have a Google Classroom school account, so I need videos that show new features in action with students.

Flipped Classroom Tutorials videos are useful because they show Gemini inside the actual Classroom environment. They help teachers see where the tools appear, how they might support lesson planning or assignment creation, and whether the workflow feels natural.

Here’s one on a feature I knew little about:

Guided Learning (Aprendizaje Guiado)

I am also presenting Guided Learning, or Aprendizaje Guiado in Spanish. Google describes it as offering step-by-step explanations that teach both the “why” and the “how” behind complex topics. In Spanish, Google describes Aprendizaje Guiado as a way to break down concepts and problems through step-by-step guides that support deeper understanding.

I like to focus on Guided Learning (and Gems as AI tutors) because it’s one of the tools most directly connected to student learning. It addresses a pivotal question: Will students use AI to avoid thinking, or can AI be structured to support better thinking?

I like how Jaime Profe / IA para profesores denonstrates demonstrates Aprendizaje Guiado in this video through four separate examples. (You can auto-dub it in English if needed). It helps me think about what I want I want to show teachers and how I want to show it.

What I look for in a useful AI teaching video

The most helpful videos show the tool being used for a recognizable teaching task. That might be creating a rubric, preparing a differentiated text, supporting a student through a concept, or building a resource that can actually be shared with a class.

I also want to know the limitations. Does the output need substantial editing? Is the workflow confusing? Does the feature require a particular account? Does it work in Spanish?

So, I especially appreciate videos from classroom teachers, instructional technology coaches, and educational organizations. They usually understand teaching better than technology companies do, and they are more likely to ask the questions teachers actually ask.

English-language creators and sources I’m watching

Eric Curts / Control Alt Achieve
Eric Curts has long been one of the most reliable Google-for-education tutorial creators. His strength is translating Google tools into classroom-ready ideas. Even when a video is not about the newest release, his work often helps teachers understand how to use a tool instructionally. He publishes a Google update monthly.

Why I recommend it: He is practical, classroom-aware, and consistently teacher-friendly.

Best for: Google tutorials, classroom workflows, AI-supported teaching strategies, instructional examples.

Flipped Classroom Tutorials
This is a good place to look for practical Google and Gemini demonstrations. The videos tend to show the tool on screen and walk through what teachers can do with it. For Gemini in Google Classroom, this kind of channel is especially useful because it can show where the AI tools appear, how teachers access them, and what the workflow looks like from a classroom perspective.

Why I recommend it: It is practical, teacher-facing, and usually connected to real instructional tasks rather than abstract AI blather.

Best for: Gemini in Classroom, Google Classroom workflows, AI-supported lesson planning, teacher productivity.

AppsEDU Google Workspace Tech Experts
AppsEDU is useful for tracking Google Workspace changes and understanding how new features fit into the larger Google ecosystem. Their update-style videos are particularly helpful when I need to know what has changed across Workspace and how it may affect educators or school technology teams.

Why I recommend it: It is especially helpful for tech coaches, instructional technology leaders, and administrators who need to understand the broader Workspace context.

Best for: Google Workspace updates, Gemini rollouts, access issues, admin and implementation context.

Google for Education
Google for Education should be part of any serious search for Gemini tutorials, even if it should not be the only source. Official Google videos are useful for seeing how Google describes the feature, what the intended use is, and how the product is being positioned for schools.

Why I recommend it: It is authoritative and useful for official framing, even when the classroom application needs to be interpreted carefully.

Best for: Official feature explanations, product overviews, access information, polished demonstrations.

John R. Sowash
John Sowash is a strong source for Google Workspace for Education training. His videos are especially relevant for schools already committed to Google tools and Chromebooks. When I am looking for clear, educator-oriented explanations of Gemini and Google Workspace, his channel is worth checking.

Why I recommend it: He understands the Google education environment and speaks to educators and school technology staff, not just general productivity users.

Best for: Gemini for teachers, Google Workspace training, Google Classroom and Chromebook contexts.

Navigating the gaps

Google videos are often promotional. Teachers videos are often sparse . Ed-org videos are often abstract. General AI tool videos are aimed at businesses.

Finding the right video can be challenging.

As I push forward, I am avoiding videos that say “AI will transform learning” and looking for videos that show what a teacher can actually do.

In my mind, the best tutorials meet several criteria:

• They show the tool in action.

• They connect the feature to planning, instruction, feedback, differentiation, assessment, or student practice.

• They identify whether the workflow is teacher-facing, student-facing, or both.

• They show the output, not just the prompt.

• They acknowledge limitations.

• They help a teacher imagine a realistic next step.

The AI education space does not suffer from a shortage of tool announcements. What teachers need are visible, practical examples. Good videos help us move from update to application, or from what Google says is possible to what a teacher might actually try on Monday morning.

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

Google Workspace and Gemini Updates

Google announced a wave of Gemini and Workspace updates with direct implications for educators and school IT. New features include Workspace Intelligence with admin controls, stronger Gemini support in Docs, AI-powered file search and overviews in Drive, broader browser-based AI help in Chrome, and new Gemini app capabilities such as file generation and NotebookLM integration. Together, the updates point to a future in which AI is more deeply embedded across everyday Google workflows, while access, controls, and feature availability still vary by product tier and account type. For educators, the real question becomes how much embedded AI support should shape everyday teaching and school operations.

• “Introducing Workspace Intelligence, with admin controls” - Google Workspace Updates

• “Gemini Drops: New updates to the Gemini app, April 2026” - Google blog

• “New Gemini capabilities in Google Docs help you go from blank page to brilliance”- Google Workspace Updates

Commentary:

• “Google updates Workspace to make AI your new office intern” - TechCrunch

• “Google turns Chrome into an AI co-worker for the workplace” - TechCrunch

• “Google’s new Deep Research and Deep Research Max agents can search the web and your private data” - VentureBeat

Curation, Taste and the Writing Classroom

The column describes a first-year writing assignment called the “Influences Project,” in which students trace three generations of artistic influence from a work they love. The instructor says AI helps students find possible ancestor texts and historical context topics, making a large, exploratory research project more manageable. But the deeper work remains human: journaling, conferencing, selection, contextualization, and writing. She presents AI as useful for widening inquiry and reducing search friction, not for replacing taste, voice, curiosity, or authorship.

https://www.insidehighered.com/opinion/columns/just-visiting/2026/04/23/using-ai-explore-influences-through-writing?

Tom’s Take: This guest post offers a concrete and helpful model of AI use that strengthens student agency rather than usurping it. The instructor argues AI can, under the right learning design, support curiosity, ownership, and more meaningful writing .

Subscribe now

Code.org replaced “Hour of Code” with “Hour of AI” and is building a new “AI Foundations” high school course.

The 74 reports on that Code.org is reworking its long-running coding initiative in response to AI. The organization has replaced its well-known “Hour of Code” with an “Hour of AI” and is developing an “AI Foundations” course for high school students due this fall. According to the report, the course will have students use AI to help build and lay out interactive websites, then combine their own code with AI-generated code to improve them. A middle school curriculum is also planned.

https://www.the74million.org/article/as-ai-rewrites-the-rules-of-coding-code-org-pushes-to-reinvent-itself/

Tom’s Take: Code.org has been one of the most influential groups shaping K–12 computing education. Its shoft to to AI-integrated instruction is an indication that schools will increasingly treat AI literacy as part of foundational computer science education.

Meta will now allow parents to see the topics their child discussed with Meta AI

Meta announced that parents using its supervision tools can now see the topics their teens discussed with Meta AI over the past week on Facebook, Messenger, and Instagram. A new “Insights” tab shows categories such as school, writing, health, entertainment, and lifestyle, along with subtopics. The feature launched in the U.S., U.K., Australia, Canada, and Brazil, with broader rollout planned. Meta also added parent conversation prompts and said it is creating an AI Wellbeing Expert Council.

https://techcrunch.com/2026/04/23/meta-will-now-allow-parents-to-see-the-topics-their-child-discussed-with-meta-ai/

Tom’s Take: Tech titans are feeling growing pressure to monitor how minirs use AI across everyday platforms. At the same time, schools and families are increasingly feeling they need AI insights and guidance that extend beyond classroom tools and into students’ daily lives.

Syracuse University Gave AI Access To 30,000+ Students and Faculty. Here’s What They Learned

Syracuse University gave Claude AI access to more than 30,000 students, faculty, and staff and developed uses ranging from test practice to course search and donor-data research. Leaders reported that AI-generated multiple-choice practice questions did not improve performance, but redesigned short-answer questioning with AI feedback raised initial exam scores by 12 points. The university also launched Claude-powered tools for course scheduling and coding, created an inclusive AI governance structure, and emphasized transparency, privacy, sustainability, and pedagogy over simple policy statements.

https://www.techlearning.com/technology/ai/syracuse-university-gave-ai-access-to-30-000-students-and-faculty-heres-what-they-learned

Tom’s Take: Note that learning gains came only after Syracuse changed the task design. And Ai feedback was a critical component. The article indicates that successful campus AI adoption depends on pedagogy, overesight, transparency, and buy-in.

Understanding Workspace Agents in higher education

The post explains Workspace Agents in ChatGPT as reusable workflows for repeatable higher education tasks, rather than one-off chats. OpenAI says eligible ChatGPT Edu workspaces can use them in research preview, with admin-controlled permissions and human review. The article gives campus examples such as course-change routing, faculty AI-support intake, weekly student-services summaries, grant-readiness preparation, campus communications, IT software requests, program review packets, onboarding, and accreditation evidence collection. It emphasizes starting with low-risk, structured workflows and keeping humans in the loop.

ChatGPT for Education

Understanding Workspace Agents in higher education

This article provides an overview of what workspace agents are and how they can support useful, repeatable workflows in education settings. For detailed guidance on how to set up workspace agents, please review this Help Center article…

Read more

4 months ago · 67 likes · 3 comments

3 Mistakes District Leaders Say They’ve Made in Rolling Out AI

District leaders told EdWeek Market Brief they made three main mistakes in rolling out AI: assuming all leaders were ready, failing to engage labor unions and parents early enough, and spending too much time on policy before building school-level understanding and capacity. Examples came from New York City District 19, Gem Prep, San Diego Unified, and Broward County. The article describes districts slowing rollouts, rebooting advisory structures, and appointing school AI liaisons to reconnect policy, training, and classroom practice.

https://marketbrief.edweek.org/meeting-district-needs/3-mistakes-superintendents-say-theyve-made-in-rolling-out-ai/2026/04

And More…

• Google Translate Now Lets You Practice Prounciation - TechCrunch

• What is Flint and How Can I Use It To Teach? - Tech&Learning

• What Schools Should Ask Before Buying An AI Tool - Tech&Learning

• Bringing Immersive Gaming into the Classroom - Stefan Bauschard

• What’s my line? Exploring how we figure out what is our line when it comes to using AI in scholarship and research - Lance Eaton

• From innovation to impact: Three ways school districts can build a sustainable AI framework - eSchoolNews

• The Secret Behind School Boards Buying Technologies - Larry Cuban

• YouTube is testing an AI-powered search feature that shows guided answers - TechCrunch

• Meet Shapes, the app bringing humans and AI into the same group chats - Tech Crunch

• After dissing Anthropic for limiting Mythos, OpenAI restricts access to Cyber, too - Tech Crunch

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

The timing was uncanny.

Just the day before, Sal Khan admitted that the AI tutor Khanmigo has been something of a “non-event” for many students. He admitted that only a relatively small percentage of students who have access to Khanmigo actually use it, and even when they do, interactions are often shallow.

In a widely commented post, Dan Meyer went strikingly further, declaring Khanmigo effectively “dead” as a symbol of the AI tutoring revolution. He did not relegate his criticisms to Khanmigo, but rather to a larger notion that AI tutors are transforming learning.

His argument, in part, is that, despite all the advantages—OpenAI access, major funding, institutional support—students simply didn’t use the tool in meaningful ways.

But linking Khanmigo’s struggles to the symbolic death of AI tutors is too sweeping—and misleading.

The Problem Isn’t AI Tutors. It’s the Model of AI Tutoring

A more accurate characterization is that there is a mismatch between a high-profile model of AI tutoring—primarily chatbot-based—and the broader integration of AI-supported learning.

Student engagement with AI tutoring depends heavily on design, learning architecture, context, and subject area. A chatbot that expects students to ask good questions assumes students already know how to leverage tutoring well. Many simply do not, or care not. A more effective AI tutor structures the conversation, frames the interaction, adapts the difficulty, prompts the learner, and creates a reason to continue.

Duolingo is an AI tutor. It’s not just the most popular language-learning app in the world; it is the world’s most widely used learning app. Millions of students engage with it daily. As I’ve explained, Duolingo’s success is built on a strong foundation of engaging instructional design, not just chatbot interactions. Its bite-sized learning chunks, gamified lessons, adaptive “Birdbrain” AI system, hilarious Stories, Avatar-driven audio conversations, and addictive streak mechanics keep language learners engaged and coming back day after day.

A few weeks back, I told you about an illuminating AI-tutoring study conducted by researchers at the University of Pennsylvania and National Taiwan University. Students using an AI-driven tutoring system significantly outperformed students in a fixed-sequence environment. Both groups used the same GenAI tool. The key difference was not access to AI, but how it was used. The AI-driven tutoring system analyzed student behavior—code revisions, interaction patterns—and adapted learning tasks accordingly. The result was measurable improvement, estimated at some several months of schooling.

A Different, “Visible” Direction: Hyperrealistic AI Tutors

If chatbot-based tutors represent one path, a different and potentially more engaging path is already here: hyperrealistic, multimodal AI tutors.

Speakology is doing something that instantly feels qualitatively different from most AI tutoring tools. Instead of a chat window, you’re in what feels like a video call with a human being. A hyperrealistic AI avatar looks at you, asks you questions, responds in real time, and keeps the conversation moving.

See Ben demonstrate it during our conversation:

Even having used it before, I was struck again by how real it feels. (And it was fun to speak some French.)

A face-to-face hyperrealistic conversational avatar is not “ChatGPT with a face.” It is a different modality altogether and one that can sustain attention, increase practice opportunities, and facilitate learning environments that are difficult (if not impossible) to implement in traditional classrooms. Speaking practice, role-play, simulated interviews, oral defenses are certainly areas where AI can expand access and frequency of practice in meaningful ways.

Why Language Learning Is a Leading Indicator

Ben and I agree that language learning, in particular, is an effective domain for AI tutoring. Speaking a language requires repeated practice, quick feedback, and a willingness to take risks. These are conditions that are difficult or impossible to sustain in a classroom with 25 students and limited time.

Ben points out that once of the best uses of Speakology is short, frequent speaking interactions—often just five to ten minutes—without the pressure of speaking in front of peers. Ben related that students in a University of Wisconsin program who used Speakology weekly in a beginner French course outperformed students in a higher-level course on a standardized speaking assessment.

That’s impressive. If you’ve ever tried to learn a new language you know that jumping just one level can take hundreds of hours of practice.

Progress in speaking ability is typically slow and uneven. And language learners frequently plateau. (Raising my hand.) Gains of that size suggest that increased practice frequency, combined with targeted feedback, can accelerate development in ways that traditional structures cannot accomodate.

Part of Speakology’s effectiveness is its presonalization. The system dynamically adjusts to student progress, pushing more advanced learners while scaffolding those who struggle. It can also adapt topics based on student interests (“clamshells” vs “rainforest” as Ben illustrates) maintaining engagement while still aligning with instructional goals.

I was struck that Speakology uses multiple LLMs—seven!—to evaluate speaking performance. Different models assess language speaking across various dimensions such as communication, structure, and cultural competence. This is a more complex approach than I anticipated and suggests a move toward more nuanced, multidimensional assessment, than what educators might expect.

The Role of the Teacher: Control, Design, and Tension

Ben consistently reiterated that the teacher is central to Speakology’s tutoring system. He explained that teachers define the learning objectives, select topics, determine proficiency targets, and control how feedback is delivered. The AI executes within those parameters.

In theory, this aligns well with a teacher-led model of instruction. In practice, however, I see significant tensions—especially around assessment.

I was taken back at how extensively Speakology is used for oral exams. Yes, oral assessments are very time-consuming. Teachers must schedule individual sessions, conduct evaluations, and provide feedback—an extended commitment. Speakology can administer assessments at scale, evaluate responses using predefined criteria, and provide immediate feedback.

But, efficiency aside, this raises important questions. If teachers are not directly listening to students speak, what is lost? Is there a risk of outsourcing a core aspect of teaching? And more broadly, what is the appropriate role of AI in evaluating student performance?

If you’re using AI to grade papers, you’ve probably been wrestling with these exact questions. They all point to the need for careful, intentional use. I don’t have ready answers, but I do know that efficiency should not come at the expense of meaningful teacher engagement, where human judgment and interaction are central to learning.

Beyond Language: Expanding Possibilities

While language learning is a natural fit, Speakology’s underlying model has broader implications. Ben points out that Speakology is already being used for career readiness, interview preparation, and even applications in social studies. Teachers can upload materials, so students can engage in simulated conversations around charts, graphs, maps, spreadsheets, and other sources, broadening its appeal across K-12 subject areas.

These examples suggest a possible shift towards interactive experiences—structured, responsive, and personalized. That expansion becomes even more powerful when students are not just responding to oral prompts, but interacting with graphs, charts, maps, and other shared visual data, or even — as Ben envisions — engaging in multi-student, group conversations mediated by AI.

We are still early in this evolution. Many of these use cases are experimental. But they suggest that the future of AI tutoring will not be defined by static question-and-answer interactions, but by dynamic, multimodal learning environments.

A More Accurate Framing of the Moment

It is tempting to interpret the underwhelming response to Khanmigo as evidence that AI tutoring has failed. But that conclusion overlooks the diversity and vitality of the broader landscape.

AI tutoring is not a single product category. It cannot be defined solely as a chatbot. It is a evolving ecosystem — an exciting one at that — spanning adaptive systems, conversational agents, and immersive instructional environments. Some approaches are bound to falter. Others will gather steam. We are in the early stages of AI tutoring. What matters is not whether an early implementation met inflated expectations, but rather whether the underlying trajectory is moving toward more effective forms of learning.

From that perspective, the story is not about failure, but of transition.

—

Find AI-in-education tools, news, how-tos, consults and more at tomdaccord.com

New in NotebookLM: Editable Slide Decks and Infographic styles

Google announced several NotebookLM updates: users can now revise generated slide decks, export them as PPTX, upload EPUB files, save private chat history across sessions, and create artifacts such as audio overviews, video overviews, and reports directly from chat. NotebookLM also adds cinematic video overviews, ten selectable infographic styles, and better flashcards and quizzes with saved progress, shuffling, deletions, and retry options. Most features are broadly available now; cinematic video overviews are limited to certain paid plans.

https://workspaceupdates.googleblog.com/2026/03/new-ways-to-customize-and-interact-with-your-content-in-NotebookLM.html

What Studies Say About How AI Affects Your Brain

Alberto Romero reviews more than 30 studies on AI’s effects on cognition, learning, and psychology and argues that the evidence points to an important distinction: AI’s effects depend heavily on the design of the intervention and on how the tool is used. He highlights research showing that some forms of AI use can boost immediate performance while weakening the mental processes that build durable competence, while other, more deliberately designed uses can genuinely improve learning. The same underlying technology can produce very different cognitive outcomes.

The Algorithmic Bridge

What the Studies Say About How AI Affects Your Brain: A (Very Big) Compilation

Hey, Alberto here! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that…

Read more

5 months ago · 127 likes · 32 comments · Alberto Romero

Tom’s Take: “The debate about “should we use AI?” is much less important than the question of how the interventions are designed and how the technology is used.” Bingo! The public is fixated on the binary that AI is “good” or “bad,” but technology can not be separated from the hands of the user.

AI + Math Learning. How to Solve a New Problem (Video)

Education Week has released a new video in which it identifies effective AI uses in math education, but states that both educators and students will need to approach these tools with informed skepticism.

https://www.edweek.org/technology/ai-math-learning-how-to-solve-a-new-problem/2026/04

Grammarly Authorship: I Tested The Recently Updated AI Writing Tool

Tech & Learning says Grammarly Authorship is a more useful tool than traditional AI detectors because it creates a writer-controlled history showing what was typed, pasted, or AI-generated and how long the writing took. In testing, it correctly identified typed text, pasted quotations, and inserted AI text. The reviewer sees possible value for citation awareness, plagiarism conversations, and student self-checking, but stresses that it is not foolproof, raises privacy concerns, and can still be gamed by determined students.

https://www.techlearning.com/news/grammarly-authorship-i-tested-the-new-ai-and-plagiarism-tool

Tom’s Take: It’s important to distinguish between AI detection and writng process documentation. Emphasis should always be on the latter. Grammarly Authorship does not “solve” AI cheating, but it nudges us toward a more process-based approach to writing authenticity. As such, it may encourage rich conversations about drafting, citation, and authorship.

Converting a traditionally graded course to an alternative assessment course

Jason Gulya argues that AI strengthens the case for alternative assessment because traditional grading encourages students to trade products for grades rather than engage in meaningful learning. He describes redesigning an online Berkeley film course around four core skills—collaborating, communicating, analyzing, and being creative—using a hybrid complete/incomplete and try-again system instead of conventional percentages. His goal is to shift attention from task completion to skill development and to rethink course incentive structures in ways that make uncritical AI offloading less attractive.

The AI Edventure

AI and Alternative Assessment

I’ve been saying it for well over two years at this point. One of the biggest obstacles to adapting to Artificial Intelligence (generative, agentic, predictive, etc.) is the transactional model of education. This model incentives trade-ins…

Read more

4 months ago · 24 likes · 2 comments · Jason Gulya

Tom’s Take: Guyla connects AI directly to assessment design, arguing that “transactional” grading propels students to use AI. This is a concrete, helpful example of how one instructor is redesigning feedback and grading to ensure that process and skill growth are central to assessment.

It Is a Mistake to Ban Devices in Schools (Richard Culatta)

ISTE CEO Richard Culatta argues that state efforts to ban or sharply restrict devices in schools are misguided because they conflate distracting entertainment media with research-based educational technology. He criticizes bills in states such as Missouri, Tennessee, Kansas, Virginia, and Utah for emphasizing paper-based instruction or screen-time caps instead of instructional quality. Culatta says effective edtech can support engagement, personalization, data-informed teaching, accessibility for students with disabilities, and digital-literacy development, while blanket restrictions risk undermining both learning and future workforce readiness

https://larrycuban.wordpress.com/2026/04/24/some-states-are-banning-devices-in-schools-thats-a-huge-mistake-richard-culatta/

Related:

Tom’s Take: Culatta pushes back against the rising anti-screen, anti-tech sentiment now permeating school policy. Culatta urges that schools not ignore evidence-based educational tech-use. Too often, the debate is framed around raw screen-time totals and too little around quality and instructional value Schools Play Game of Media Literacy Catch-Up as AI Use Rises

Recent video of note:

And More

• Los Angeles schools set limits on classroom screen time - Reuters

• The Good, Bad, and Ugly of Screen Time: An Explainer - Education Week

• New CoSN Report Offers Guidance on Responsible Technology Use in Schools - THE Journal

• Hot Takes on AI and Education - Education Next

• Students to get virtual tutors on master’s courses taught by AI - The Times

• Has AI As A Writing Partner Been Oversold? - Tech&Learning

• The White House says all 50 states joined the Presidential AI Challenge, and nearly 3,000 schools participated in a White House AI workshop - White House

• OpenAI releases GPT-5.5, bringing company one step closer to an AI ‘super app’ - Tech Crunch

• Anthropic released Claude Opus 4.7, positioning it as a more capable generally available model for complex reasoning and agentic coding - Anthropic

• ChatGPT introduces Images 2.0 - OpenAI

• ChatGPT’s new Images 2.0 model is surprisingly good at generating text - Tech Crunch

• Turn your best AI prompts into one-click tools in Chrome - Google

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

In preparing for the webinar, I came across several tools recommended by Eric Curts, a leading voice in edtech integration. Eric has spent decades helping educators make effective use of digital tools, and his Control Alt Achieve blog remains one of the most useful practical resources in the field. The tools featured below don’t “solve” AI cheating, but they help teachers establish better structures around it.

Defining the Problem: What Counts as AI Cheating?

Before getting to the tools, however, two foundational issues have to be addressed. The first is definition. The second is communication.

The first question is deceptively simple: What are acceptable uses of AI when writing a paper? Educators universally agree that having AI write a paper and then submitting it as one’s own is unacceptable. But the moment we move away from that obvious scenario, the terrain becomes noticeably rocky. Is it acceptable for a student to use AI to brainstorm possible angles for an argument? Okay if AI suggests a rough outline? How about AI summarizing a source or two? Can it suggest an alternative sentence or two? What about a metaphor to punch up an argument? Different teachers, departments, and schools will answer those questions differently, and their answers may shift depending on the assignment, grade level, disciplinary norms, and instructional goals.

As Curts points out, it’s useful to think in terms of a continuum rather than a binary. On one end are clearly acceptable uses, such as brainstorming, language support, or summarizing a source. On the other end is clear substitution, where the machine is doing the intellectual heavy lifting that the student is supposed to do. In between lies a wide gray area.

What we have to focus on is not whether AI appeared somewhere in the process, but whether its use displaced the student’s own thinking, decision-making, and authorship. That is the pedagogical question teachers need to address before crafting an AI policy statement for an assignment.

(For a broader ethical framework for thinking through these issues, UNESCO’s guidance on generative AI in education and research remains a useful compass.)

Share Clear Expectations

Once a teacher has thought through what is and is not acceptable, the next step is to make those expectations visible to students. A surprisingly effective way to do that is via a traffic-light model. Its value is clearly not sophistication but clarity. Students instantly understand the symbolism. But the teacher needs to fill in the particulars for each light.

This kind of concise visual guide can help move AI policy out of vague multi-page statements and into something students can actually use when they are doing their work.

Traffic Lite

One tool that operationalizes this idea is Traffic Lite. Embedded directly into Google Classroom, Traffic Lite lets teachers attach AI-use guidance to assignments so expectations aren't buried in a separate policy document. Instead, students see in real time which kinds of AI use are permitted.

Confusion is one of the great purveyors of misuse, and many students will exploit ambiguity if teachers leave it unaddressed. Traffic Lite offers teachers customizable AI-usage policies, multiple color-scale options, and a free teacher-facing version. At the administrative level, it also allows school leaders to create and manage broader policy structures across groups or departments, which could be especially useful for schools trying to move beyond disparate classroom-by-classroom norms.

EduGems and AI Expectations

Another promising option in this “clear expectations” category is EduGems, Eric Curts’s collection of pre-made Gemini Gems for educators. EduGems is essentially a curated library of reusable Gemini prompt-tools organized by instructional purpose, from lesson planning and assessment to AI integration and student supports.

The most useful for our purposes is AI Expectations. This Gem helps teachers establish “clear, pedagogy-driven AI boundaries for student assignments” by guiding them through a four-level traffic light model tied to the actual learning objective being assessed.

What makes AI Expectations especially valuable is that it does more than generate policy language. It walks the teacher through a collaborative planning process inside Gemini. The teacher begins by entering the assignment topic, the core learning objective or skill being measured, and the students’ grade level. From there, Gemini helps the teacher determine the appropriate traffic-light level, think through edge cases, and produce a student-facing AI policy statement that can also include a checklist and citation requirements.

In other words, AI Expectations helps turn broad concerns about AI misuse into assignment-specific guidance. Teachers know that “use AI responsibly” is too vague to be useful, and this Gem offers a practical way to build clearer boundaries before the assignment ever reaches students.

Surveying the Writing Process, Not Just the Product

If the first two tools are about setting expectations, the third is about monitoring process. Many teachers already rely on Google Docs version history, which teachers can consult during the writing process and can certainly help reveal suspicious bursts of pasted text or strangely quick timelines. But version history has limits. It tends to give you snapshots, not a rich reconstruction of how the writing unfolded.

That is where playback-style tools become especially interesting. These tools do not merely inspect the final product; they help teachers look at how the product came into being. That shift in emphasis is fundamental, because academic integrity in the age of AI increasingly depends on process visibility.

Process Feedback for Google Docs

One of the more intriguing playback tools is Process Feedback for Google Docs. This extension reconstructs a document’s writing journey from Google’s underlying edit history and turns that activity into a more detailed report. It can generate visualizations of typing fluency, show time spent on different paragraphs, indicate revision ratios, detect paste events, and even track contributions from multiple authors in collaborative documents.

Unlike Google Docs’ standard version history, Process Feedback is built for analysis. It also handles data in a privacy-conscious manner by running locally in the browser unless a user chooses to share a report link. In other words, it is less a detector than an analytics tool. For teachers or students who want a more scientific look at writing behavior, this tool is important.

Speaking of “detector,” I’ve intentionally avoided focusing on AI detectors. Not because they are insignificxant, but because AI detectors pull teacher’s attention toward the final product rather than the full process. Tools that make the writing process visible are ultimately more aligned with that goal than tools that attempt to judge the final output after the fact.

A Practical Way Forward

What ties these tools together is that they move us toward better teaching design. None of them can answer deeper pedagogical questions about cheating. Teachers still need to decide what kind of thinking an assignment is meant to propel, what kind of assistance is acceptable, and how much independence students are expected to demonstrate. They must think through how their assignments are activating active thinking. But once those questions are answered, tools like Traffic Lite, Edu Gems, and Process Feedback can make expectations clearer and student process more visible.

To me, this is the practical path forward. In an AI-saturated environment, schools cannot rely on bans, suspicion, or detectors. They need clearer definitions, more transparent communication, and better ways to portray writing as a process rather than a finished artifact. The goal of a writing teacher is not to catch misuse. It is to create conditions in which students understand the difference between using AI to support their thinking and using it to replace it. That is a hard task. But it is also worth doing.

Find AI-in-education tools, news, how-tos, consults and more at tomdaccord.com

Khan Academy, TED, and ETS announce the Khan TED Institute.

Khan Academy, TED, and ETS are launching the Khan TED Institute, an online college built for the AI era. Organizers said they want applications to open in 12 to 18 months, keep total cost around $10,000, and begin with a bachelor’s degree in applied AI. The competency-based program will emphasize core academic knowledge, applied AI skills such as app and agent building, and human skills like communication and leadership. Google, Accenture, McKinsey, Bain, and Replit are listed as launch partners.

https://www.axios.com/2026/04/14/khan-academy-ted-ets-institute-college

Commentary: The New Khan TED Institute: Reimagining Higher Education Without It

Jeppe Klitgaard Stricker criticizes the new Khan TED Institute as a corporate-led attempt to “reimagine” higher education without meaningful involvement from universities or higher education scholars. He argues that its employer-aligned framing narrows education to workforce preparation, while ETS’s role raises concerns about who defines and measures learning.

The Future of Higher Education

The New Khan TED Institute: Reimagining Higher Education Without It

Almost two years ago, I wrote about a number of fundamental problems with using a chatbot like Khan Academy’s Khanmigo in higher education…

Read more

5 months ago · 16 likes · 2 comments · Jeppe Klitgaard Stricker

Join me and my featured guest, Ben Altschuler, Co-Founder and CEO of Speakology AI, for a fascinating discussion about the exciting transformation of AI tutoring systems. We’ll examine visible breakthroughs in AI tutoring via hyperrealism — more natural interaction, stronger presence, and multimodal learning experiences — as well as invisible breakthroughs in how AI tutoring systems scaffold challenges and move learners forward.

Register here for this free webinar.

Stanford’s HAI 2026 AI Index Report

The Stanford Human-Centered Artificial Intelligence 2026 AI Index Report says AI capabilities are still accelerating, with frontier models improving rapidly, adoption spreading fast, and education struggling to keep up. The report highlights closing U.S.-China model performance gaps, rising AI incidents, surging investment, and growing national competition over infrastructure and policy. In education, it says over 80% of U.S. high school and college students now use AI for school tasks, while school policies remain limited and often unclear.

https://hai.stanford.edu/ai-index/2026-ai-index-report

Commentary: Want to understand the current state of AI? Check out these charts

MIT Technology Review’s overview of Stanford’s 2026 AI Index says AI is advancing quickly while institutions struggle to keep pace. Top models keep improving, adoption is spreading rapidly, and companies are generating revenue fast, though infrastructure costs are enormous.

https://www.technologyreview.com/2026/04/13/1135675/want-to-understand-the-current-state-of-ai-check-out-these-charts/

How a Chinese startup is replacing teachers

In this interview, Joleen Liang, co-founder of Squirrel AI, an education startup from China, argues that AI will replace some traditional teaching functions while shifting educators into roles such as instructor, supervisor, and data analyst. She describes Squirrel’s personalized learning system, which uses extensive behavior data, adaptive content, and human monitors in physical learning centers. She says the company is expanding in North America, pursuing international growth, and eventually envisions AI tutors or robots that guide students academically while still encouraging offline human interaction and hybrid schooling.

Commentary: An illustrated guide to resisting “AI is inevitable” in education

Benjamin Riley’s illustrated guide argues against the idea that AI in education is inevitable. Drawing on recent research, student activism, school-tech backlash, skepticism about Khanmigo, and criticism of AI-heavy models like AlphaSchool, he presents a toolbox for resisting AI boosterism. The piece emphasizes cognitive surrender, weak causal evidence for educational benefits, and the narrowing of education to efficiency and automation.

Cognitive Resonance

An illustrated guide to resisting "AI is inevitable" in education

1. Ask the AI-in-education enthusiast to clarify their premise…

Read more

5 months ago · 108 likes · 6 comments · Benjamin Riley

Why Sal Khan’s AI revolution hasn’t happened yet, according to Sal Khan

Chalkbeat reports that Sal Khan now says AI has not produced the learning revolution he once predicted. Many students did not use Khanmigo much, struggled to ask good questions, or found the tool frustrating. Some teachers stopped using it, while administrators remained more positive. Khan Academy is responding by embedding Khanmigo more directly into practice activities rather than expecting students to seek it out. Khan still sees AI as useful, but now argues that the bigger lever for improvement is investing in human systems.

https://www.chalkbeat.org/2026/04/09/sal-khan-reflects-on-ai-in-schools-and-khanmigo/

Commentary: Sal Khan Still Clueless

Peter Greene interprets the Chalkbeat report as a partial admission that Khanmigo has not delivered on its promises, but argues Khan still fails to grapple with the deeper reasons. His commentary pushes harder on the idea that the limits are not just implementation problems but misunderstandings about motivation, learning, and what students actually need from teachers and schools.

https://curmudgucation.blogspot.com/2026/04/sal-khan-still-clueless.html

Making Graphic Novels with NotebookLM

Scientific Figures Without the Design Skills? Illustrae Makes It Possible

Podcast: How to Teach Students to Spot What’s Real, Fake—or Deepfake

Students act as digital detectives in a 60-minute lesson from Katie Coppens and Andy Zucker, analyzing videos and websites to decide what is real, altered, or misinformation. They must justify each conclusion with evidence, not guesses. The School of Practice episode provides lesson instructions, strategies for spotting misinformation, and classroom-ready resources.

https://www.edutopia.org/podcast/how-to-teach-students-to-spot-whats-real-fake-or-deepfake

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com

If the name is new to you, Jon is an edtech legend who co-developed the Flipped Classroom model back in 2007. But what makes Jon’s work vital right now isn’t so much his past, but rather his present. After years of traveling the world as a consultant, Jon returned to the classroom in 2019. He is currently in the trenches teaching high school physics and geology in Houston, Texas.

As such, Jon is watching the “AI revolution” unfold in real time at the desks of actual teenagers.

And it has spurred him to action.

The Great Bifurcation

Jon argues we are at a critical juncture, The Great Bifurcation. He sees AI propelling students in two opposite directions. At one end are the “thrivers,” students who use AI to enhance, deepen, and extend their intellectual work, and at the other end are “stupified” students who use AI as a cognitive crutch, effectively bypassing the neural development (the “milein sheets” and “synapses”) required for true learning.

Jon’s framework, the Mastery Flip, is his strategic response to this divide. It builds on his original model but is redesigned for an era in which the “Easy Button” is always in a student’s grasp.

I'm not going to give away what's in the video, but I will say it's striking to watch a primary architect of the flipped classroom go back to the drawing board and ask, from scratch, what a classroom actually needs to do now. The answer he's arrived at is more surprising — and more old-fashioned at times — than you might expect.

Here is the full interview:

Reactions to The Mastery Flip

The AI Engines that Jon depicts is a fascinating way to increase student engagement and reimagines the pre-classroom activity phase in ways that go well beyond recording a lecture. I also appreciate how Analog Roots, the in-class counterpart, is laser-focused on a “high-brain” approach. Mastery Viva, the human check, is the most consequential assessment strategy in the framework and the heart of the human element.

The Mastery Flip is a great foray into strategic pedagogical balance, providing a thoughtful "middle way" through the polarization of the AI-in-education era. An example is Analog Roots, the decision to put student laptops away to protect the "productive struggle" that actually builds neural pathways. Another is the Mastery Viva, in which Bergmann positions the human voice as the ultimate truth teller. In all, Mastery Flip a thoughtful framework that leverages AI as a sophisticated on-ramp to focus on the high-brain, high-touch work that has always defined deep learning.

Jon is clearly a thoughtful and experienced practitioner, and I have enormous respect for what he’s built here.

Yet, I must say, I felt Jon was framing the AI challenge in binary terms. For instance: students who use AI as a crutch and students who use it to flourish. I’ve talked to colleagues who speak of students who use AI productively for some tasks and unproductively for others, depending on context. There’s some messiness here being overlooked.

Clearly, Jon’s claim that teachers can no longer send cognitively demanding work home is pretty bold. He’s certainly not wrong that AI puts the old homework model on very shaky ground. But his is a blanket statement, and not all educators (including this one) accept it as a fait accompli. I see colleagues designing assignments that require a documented process, and/or a personal student voice, and their approaches merit consideration.

Finally, I love the Mastery Viva but completing twenty five or more brief verbal assessments per class period is a significant ask on a teacher’s time and energy. Jon clearly loves talking with students and he groups them if needed to reach all. But it’s ironic that he recommends AI audio checks as a fallback if a teacher can’t reach all students. If the goal is a “human check,” using AI as the checker creates an incongruous loop.

I don’t think these tensions undercut the framework. I think Jon’s work is plenty solid. Ultimately, it’s designed to provoke us to act. And that we must.

Conclusion

As I listened to Jon, I was struck by how hard he is working to protect the conditions under which students actually learn. That goal is nothing new, but what is strikingly different now is that the threats to those conditions are coming blisteringly fast, and are much harder to see.

In all, Jon’s striving to ensure that the student brain remains the central part of the equation, while all of us try to figure out what “effective” AI integration actually looks like. That feels like exactly the right instinct for this moment.

Find AI-in-education tools, news, how-tos, consults, and more at tomdaccord.com