In April, a calculus teacher in Connecticut named Bill Fallon gave his twenty-one seniors an assignment that didn’t involve calculating anything. They had to explain what a derivative actually means to a classmate named “Devon,” who kept almost getting it and then slipping back.

Devon is a bot. He’s seventeen and in AP Calculus AB, with one specific hang-up. He knows the derivative gives you the slope. But a curve doesn’t have a single slope, since it’s steeper in some places and flatter in others. So the slope of what?

Bill and I built him over a handful of conversations, working out what the character should be stuck on and where the whole thing should sit in the unit. He ran it in his classroom, we surveyed his students, and afterward the two of us went through the transcripts together and talked about what worked and what didn’t. The rest of this is that story.

Before I go any further, let me acknowledge that I have a stake in this. I’m building a platform that makes these simulations easy for teachers to run, we launch next month, and I’d obviously prefer you finish this thinking they work. So please, read accordingly.

I picked a math class on purpose. When I describe using the approach of conversational simulations to combat AI cheating and to build durable skills, the response I usually get is that it sounds like an English department exercise. It’s dialogic, usually involves argumentation or exposition skills, and leans heavily on a student’s grasp of language. But it expands beyond that. Math is the hardest case to make, which is why I engaged in the experiment with Bill.

None of this is new, pedagogically speaking. In 2009 a Stanford group published work on what they named the protégé effect, built around a program called Betty’s Brain. Students who believed they were teaching a computer agent spent more time on the material and learned more than students studying for themselves, and the effect was largest among the lower-achieving kids.

In June 2023, Ethan and Lilach Mollick laid out seven ways to assign AI in a classroom, and one of the seven was AI-student, where the model plays the learner and the student does the teaching.

What I’ve tried to add to the process is targeted resistance, a record, and an evaluation mechanism. The character – in this case, “Devon” – pushes back on explanations they can’t use. The conversation leaves an artifact that a teacher can read and evaluate for understanding afterward.

What Devon does

Devon opens the same way every time:

Then he waits.

If you’d like to take “Devon” for a test drive, here he is: [Sim Link]. It’s supposed to be annoying, so notice the moment you want to quit. That moment is most of what I’ve been studying the last few years.

Your chats are anonymous, no personal information is captured, and it ain’t flashy, so try and leverage your imagination rather than focus on the tech stack.

Subscribe now

How you make a “Devon”

When people hear AI role-play, they assume that all you have to do is type act like a confused student into a chat window and hit go.

The truth is; you can. But in my experience, it falls apart quickly, because the model capitulates. Your learner says something half-formed, the character says oh, that makes sense now, and it ends in a compliment. Most of the design is spent stopping the model from being nice. The other half is about teaching it how to react when the student makes their attempt at teaching.

Bill’s half of this was knowing what a calculus student actually gets stuck on. Mine was building the character with enough depth to make the conversation useful. I just finished an MFA in creative writing, where I spent a year running these simulations on myself and keeping a diary about what happened to my own mind. That paper is out for review now. But it’s also how AI Friction Labs was born.

When designing these characters, the first requirement is that the character has to want something – it can’t just be something. In this case, Devon wants to understand what the derivative describes. He has a test coming, and the student that enters the simulation is the gatekeeper to that understanding. Everything else follows from that tension, including the frustration when a student repeats the same explanation without adjusting, adding, or deleting. Devon’s voice – so wait, okay but, no hold on – acts as a set of immersive clues.

Then come the rules, which have to leave him somewhere to move. If a student’s answer is delivered as jargon with nothing underneath it, Devon says he doesn’t know what that means. He asks one question per turn and then stops. He never praises (unlike the Khanmigos of the world). Underneath all of it sits a ladder of five stages of understanding with a ceiling he isn’t allowed to pass. In other words, Devon can get close but can’t finish the thought himself, because the moment he can, the student is off the hook.

He is also built to slip back. A student explains the tangent line, Devon says he’s got it, and two turns later the confusion can resurface. That’s on purpose, and it’s meant to teach that explaining something to someone else is never a one-off. It’s also the part students hate the most, which I’ll get to.

Here’s the full build, free to take: [Link to System Prompt]. Run it as a CustomGPT or Gemini Gem and you’ll get something that half works. The character drifts after a while and there’s no record of what happened afterward. What we’re building are mechanisms to view that artifact, evaluate it easily, and engage in a dialogue with your student about what we as human beings can learn from these experiences.

Two students, same phrase

Twenty-one students ran the simulation twenty-nine times over one week. Every transcript was scored by an AI evaluation that Bill and I then read back against the conversations themselves. Scores averaged 82 out of 100 and ran from 23 to 96. Five students went back for a second attempt without being told to. All five improved, by an average of 33.6 points.

Two of the transcripts, three days apart, are worth reading side by side – a version of what I call “Comparative Transcript Analysis.” Both students used the same phrase, telling Devon the derivative gives you the slope of an “imaginative line” at a point.

Devon didn’t take it from either of them. I hear those words. But what even is a tangent line, really? And how does a curve, which is all bendy, have an “exact steepness” at just one point?

The first student defended the phrase. Then defended it again. Then, cornered, told Devon it was impossible to find the exact slope and that any line works as a tangent. Devon lost the thread, and the session ended with him worse off than he started. He scored 50, with 25 out of 30 for persistence, since he never stopped trying, and 10 out of 30 for explanation.

The second student dropped the “imaginative line” phrase. Devon then rejected “instantaneous rate of change,” then “limits.” So the student stopped using them and started talking about direction, about where the point was heading. It took him forty messages, but he got Devon to a point of understanding and scored a 93 on the AI evaluation.

Worth noting; each student used the same “imaginative line” phrase at turn three, but ended up finishing forty-three points apart. One was rewarded for persistence, patience, and creative problem-solving. That’s a crucial point about this method. It won’t just give you a record of what they know about the subject, but also how they behave when resistance in that field is presented.

In the case of the successful student, they didn’t stick with the same approach, they found a new one. They saw Devon’s confusion as information, whereas others saw the confusion as a reason to stop. The AI wasn’t working, they thought. But that type of failure creates a different kind of learning experience that, I believe, focuses on human development alongside the acquisition of academic content.

Bill put it this way:

“Things are going to frustrate you in life. People are going to frustrate you. This bot is frustrating you. How do you work around that and make it so that you are productive and helping the person as best you can? That is a whole other level of thought and learning.”

Share

What I got wrong

One of the things we learned from this beta test was that one rep wasn’t enough. Students needed feedback and a chance to try again. Our mechanisms weren’t built for that process at the time, but we’ve learned from it and are working on it as we speak.

That concept reminds me of the popular “Ungrading” approach where students receive unlimited attempts to show mastery. It also reminds me of Jason Gulya’s “Try-Again Classroom” concept, a salve for modern times if there ever was one.

Another important note relates to the AI scoring mechanism. Evaluating a conversation isn’t objective. Our AI scorer can never be 100% right. But that’s where the educator comes in. A cross-referencing approach can lead to a faster, more viable route to “grading the chat” than hand-grading (which I have done before and is a doozy.) Bill’s take was that the evaluation scores were not only useful but largely accurate, for what it’s worth.

And, to repeat the point, it allows you to choose whether you are giving math feedback or human feedback. Do you want to correct the language they use to describe the tangent line? Or do you want to teach them about resilience, patience, persistence, and a growth mindset? Either are available, and the transcript is the avenue by which those teaching points are delivered.

If you’re interested in learning more, we are hosting a free virtual launch event on August 19th. You can register for it here - we hope to see you there: [AI Friction Labs Faculty Demo Event Registration]

In the meantime, let’s keep looking for creative ways to redesign assessments – not just so that we can end-around AI cheating (which is important), but also so that we can test students on the types of skills they will have to deploy in their day-to-day lives when they leave our classrooms.

A Note on the Epigraph

I went down a bit of a math rabbit hole over the last few weeks. I’ve got some more content coming on the wave of mathematical conjectures that have either been solved or disproven by AI recently.

But before that, I was caught by this video from EO Studio sharing Ken Ono’s thoughts on AI’s impact on intelligence, mathematics, and the economy.

To hear someone with so much experience in the field of mathematics and technology speak this calmly about AI and its impact on our world put my mind at ease, if only for a moment. I recommend giving it a watch for yourself, but also consider sharing with students. Here’s the full quote from the epigraph from the beginning of the video:

“We live at a time where the world places so much emphasis on benchmarks…There’s a lot of anxiety. When it comes to assessing intelligence, do we honestly believe that someone who has a higher IQ score is somehow smarter? No, of course not. It should people at ease. If you have to live up to the standards set by someone else, then you are not living for yourself. You’re not giving yourself credit. I consider that toxic.”

Until next time.

The free virtual launch event is Wednesday, August 19th at 1pm EST. We’ll demo the platform and you can spar with a friction bot yourself, which I’d recommend, since it’s a different thing on the receiving end. Register here: [AI Friction Labs Faculty Demo Event Registration]

Share

Subscribe now

I’ve been writing for a while about academic role-playing exercises and conversational simulations that are designed to push back. I call them friction bots. When I talk to other educators about their value, especially professors, I often get the question, “Ok, but what does the research say?”

Stands to reason. What good is a method without empirical data to go along with it? So I spent some time this week rounding up some of the published studies over the last few years that tracked the efficacy of LLM-based role-playing simulations as a learning tool.

Before I get to sharing them, I want to acknowledge my bias. I’ve been touting this approach in one form or another for 2-3 years and now am in the process of building out a platform that aims to make all of this not just possible for teachers, but easy. So, you should read my own translation of these experiments with a grain of salt. Always consider the author.

That said, I sell this idea because I believe in it. I use this AI-powered approach in my own classroom. My students clamor for more. I don’t use it every day or every week, but I think it has the potential to change education in a positive way. So, to the extent that this resonates with you, I hope you can see that my bias comes more from a place of experimentation than unadulterated hype.

Thanks for listening. Onward we go.

1. Medical education — the virtual standardized patient RCT

Perhaps the most intuitive application of the approach is in the medical field, where research indicates that simulations can surface “clinical thinking ability” in students. In JMIR Medical Education Journal, researchers created 3-D virtual simulated patients that present system-preset internal-medicine cases in mimicked clinical scenarios.

They engaged with 60 medical students interning at a hospital. Thirty were trained on the system and engaged with it, the other thirty received traditional academic case-based training. They measured effectiveness via basic knowledge assessments and virtual system scoring.

The VSP students showed a theoretical score improvement roughly six and a half points higher than the control group (17.07 mean vs. 10.67). The authors concluded that the virtual experience reinforced foundational internal-medicine knowledge, enhanced clinical thinking ability, and improved capacity for independent work. They called the system feasible, practical, and cost-effective.

My take: This is a field where simulated patient care has long been an accepted practice. Medical education experts rely on these activities to develop “bedside manner” and an ability to take complex knowledge and, for example, explain it to a patient who can’t understand jargon. The simulations often test the doctor’s ability to ask the right questions, which stems from one’s own background knowledge in the field. But more than that, they test patience, communication skills, and empathy — the soft skills side of a profession that requires an immense amount of concrete knowledge and skill.

2. Counseling — the CARE study

The LLM simulation approach also appears to show promise in the counseling field.

In “Can LLM-Simulated Practice and Feedback Upskill Human Counselors?” researchers at Stanford University and Georgia Tech created an AI patient. The LLM simulates a client in a counseling session presenting “real-world challenges.”

The researchers put 94 novice counselors into two groups:

Group 1: Practice with the AI patient alone without receiving feedback

Group 2: Practice with the AI patient plus AI-generated feedback that offered suggested alternative responses with rationales.

They then measured behavioral performance, self-efficacy, and qualitative reflections. Their conclusions were that the practice-plus-feedback group improved in client-centered microskills, while the practice-alone group showed no improvements. Furthermore, with respect to empathy, the practice-alone group declined over time and performed significantly worse than the feedback group. And, when researchers interviewed participants, the ones who received feedback adopted a client-centered listening approach. Practice-alone participants remained solution-oriented.

My take: This study goes further than most. Instead of measuring the efficacy of an LLM-based role-playing simulation against traditional forms of instruction, the study places both groups in LLM sims and asks – “What difference does AI-generated feedback on the simulated performance itself make?”

It is not entirely controversial to say “feedback is good.” But it is somewhat controversial to say “AI-generated feedback is good.” So the question I have coming out of this study is, “why did the AI-generated feedback appear to work in this case?”

I think this is a question that is ripe for further research. AI-generated feedback on essays, for example, face reasonable detraction. Might it be possible that AI-generated feedback on a different set of skills actually is effective?

3. History — the medieval peasant simulation

Stepping out of the medical and psychiatric fields, there is growing evidence that LLM’s make it possible to simulate perspective in a way that we have not had access to in the past.

In Teaching History, Benjamin Breen of UC Santa Cruz describes integrating an LLM-enabled simulation into a one-week module on the social history of medieval peasantry. In this one, the student plays the character. Breen developed a prompt that reliably generated a randomized persona of a medieval peasant, a “playable character” the student inhabits for a day in the life, set in a randomly chosen location between Iceland and the Levant, sometime between 900 and 1400 CE.

In the first session, Breen delivered a lecture, circulated primary sources (both textual and visual), and completed a live playthrough on the classroom projector before students touched it. Students then ran the simulation themselves, with follow-on assignments including “prompt revision,” in which they altered the prompt to retell events from a different perspective.

The findings are qualitative (classroom case studies, no control group). Students described the simulations as encouraging historical empathy, the realization that the people they study were once real human beings with normal lives. Breen concluded that the simulations worked best when integrated into a larger cycle of discussion, analysis, reflection, and research, and not as one-off engagement activities.

My take: The feedback I received from our beta partners over the last six months of experimentation follows a similar pattern. The inclusion of pre-and-post reflections, accompanied by discussion and analysis, made a significant difference in the perceived efficacy on the various tests.

It also creates a portfolio of student learning artifacts that create a narrative of their own experience. When consumed together, the educator sees an arc that tells the story of their student’s journey.

4. Teacher education — AI students in a virtual classroom

One of the criticisms that many educators levy against the teacher training programs that they experienced in undergraduate or graduate level education settings is that the traditional form of study doesn’t teach you how to manage a classroom. The pre-service teacher may learn quite a bit of meaningful pedagogy and write many papers on what is effective and what is not, but how do they get prepared to manage the outlandish and disruptive comment from the student in the back-row? How might they practice keeping students on-task when the weather is nice and it’s Friday at 1:30pm? How might they receive “reps” in responding to genuine misconceptions that you didn’t plan for?

LLM-based role-playing simulations show great promise in this area.

In Education Sciences, researchers introduced TeacherGen@i, a generative AI-enhanced virtual reality simulation for pre-service teachers. The characters are the students: virtual agents that respond verbally in real time, display dynamic facial expressions, and sit in a virtual classroom where the trainee’s own course slides appear on a display. The system ran on the participants’ personal laptops.

The study was an explanatory case study with a mixed-methods design, embedded in a university teacher-preparation course. Pre-service teachers taught lessons to the AI students, and the researchers collected usability surveys and reflective journals, analyzed through thematic coding and computational linguistic analysis.

The findings suggest the simulation facilitated meaningful development of teaching competencies, specifically instructional decision-making, classroom communication, and student engagement. The authors also documented design limitations: cognitive load, user interface problems, and insufficient instructional scaffolding. They frame the work as exploratory.

Related work in the same space embedded GPT-4 student agents in a 3D Roblox classroom, where pre-service teachers practiced problem-solving lessons, with pilot studies indicating high usability and learner engagement.

My take: We’ve run teacher education simulations as well. However, we’ve focused on conversations that happen outside of the classroom.

Just last week, faculty at Portland Community College engaged with a friction bot that presents one of two difficult colleagues. One is the type to agree to everything but never actually execute the plan. Another is the type who disagrees with everything. In each case, the task aligns with the conflict point. Can you get the commitment-averse educator to firmly agree to follow through on a plan? How? In the second, how do you navigate disagreement without damaging the relationship? Can you find a center ground that is both true to what you believe but also takes into account the unique perspective of your counterpart?

One PCC faculty member described it as “hitting close to home.” The realism of these experiences at a minimum provides a space for rehearsal of difficult conversations. It’s “a safe place to fall on your face.” Wouldn’t that be nice?

5. Negotiation — ACE

No one ever taught me how to negotiate. I remember failing miserably to secure a raise in one of my first year-end reviews. A few years later, I studied negotiation methods before entering a similar meeting. I didn’t have a place to practice, but I had a plan, and that’s more than I can say for most situations. I succeeded in getting some of what I wanted, but not all. I look back and wonder; what if I had a place to practice that negotiation? In a place devoid of fear of judgment or failure? Might I have performed better in the moment?

In that vein, researchers at Columbia University built ACE (Assistant for Coaching Negotiation), presented at EMNLP 2024. The character is a negotiation chatbot that serves as the learner’s bargaining counterpart, with a coaching layer built on top of it.

To build the system, the team collected a dataset of negotiation transcripts between MBA students, trained negotiators working through realistic single-issue bargaining scenarios. With expert consultation, they designed an annotation scheme for detecting negotiation mistakes. ACE uses that scheme to identify the user’s errors and deliver targeted feedback, alongside examples written by experts.

They then tested it with 374 participants, each completing two rounds of negotiation, across three conditions: ACE’s feedback, no feedback, and an alternative feedback method. ACE significantly improved negotiation performance compared with both other conditions. Participants in the ACE condition also reported higher perceived improvement in their second negotiation, a mean of 4.23 on a five-point scale. The authors built the system because negotiation skills are traditionally taught through resource-intensive seminars that most learners never get access to.

My take: For our part, we have run pre-hire job candidate screening simulations for multiple businesses in the same regard. In one, a candidate had to negotiate with a room full of conflicted stakeholders. In another, the candidate had to defend their investment thesis to a difficult investment committee member. Negotiation-adjacent on that last one, to be sure, but a practice space for flexing the same muscles.

What’s it all mean?

This the tip of the iceberg. Let’s recap some of what we know:

1. Role-playing has been proven time and time again to be good for learning.

2. Rehearsal is good for the brain, too.

3. LLM’s are really good at mimicking human beings.

4. Human-human “simulations” are great! We shouldn’t abandon them. But what LLM’s may introduce is a way to scale the practice.

5. The advancement of LLM’s in the past few years mean that simulations can now be executed in a far more dynamic, adaptable, and immersion-inducing manner.

What we need are the right pedagogical mechanisms around the LLM simulation for this to really be useful. Beyond that, the “craft” of producing these virtual rooms, simulated conversations, and role-playing dialogues needs to advance and be codified across a wide range of educational contexts.

One more note: These are just a few of the studies out there. I worked with Claude to produce a one-sheet of relevant studies in the field. You can find it here, nothing fancy.

Furthermore, there are many educators already running these simulations informally/as experiments. Substack is filled with them, though I didn’t have time to curate and place them all inside this article. If you know of any good ones out there (or if you have one yourself), please drop them in the comments.

What I’m Reading/A Note on the Epigraph

Graham Greene’s The Power and the Glory. My uncle is a big Greene fan, but this is my first foray into his work. What I’m most enjoying is Greene’s ability to shift perspectives quickly and often without losing me. I’m only about 4-5 chapters in. Here’s how it’s going:

We begin with a nameless dentist in a raggedy Mexican town who ends up interacting with the only remaining Catholic priest in a war-torn state where all the other Catholic churches and priests have been eradicated. By Chapter Four, the dentist is gone and replaced by the lieutenant chasing him, a fellow priest who gave into the rebel’s demands and broke his celibate vows, a surreptitiously devout Catholic family in the same town, and an English sea captain who finds out his daughter has been harboring the fugitive priest without his knowledge.

Greene has me spinning like a dreidel through each of their minds, and yet I am not mad about it. I’ve tried to write in the same way. For my money it is one of the greatest challenges a writer can face.

The quote in the epigraph was chosen less for its thematic meaning and more for its style. I wonder what Strunk and White would have said about it. Actually, who am I kidding? I don’t care. I think it’s beautiful, and that’s enough for me.

Until next time.

This past winter, I began experimenting with structured conversational simulations in my Rhetoric and Inquiry class at a local New Jersey university. For my class, I called them “Friction Bots” and made it clear that the conversation they were about to enter was designed to be difficult.

I built characters and scenes for different skills and topics that we were working through. One tested their ability to identify rhetorical moves. Another tested the ability to back up their thesis. Another tested their ability to explain a theme.

The characters in each simulation pushed back through a series of branching mechanisms I designed into the back-end to make the conversation feel more realistic and adaptable.

At the time, the most important re-frame I felt I needed to get across to students was that AI could be difficult, and not easy or a shortcut, if the experience was redesigned. For many, that is now a well-known development. But at the time, I could tell my students had little perception of what I meant. To them, there was only ChatGPT.

I told them the chat would be hard. They nodded. I repeated. “No,” I said, “I mean, you’ve got a task to do, but it’s not going to be easy.” They rolled their eyes.

I handed them a “student dossier” that laid out the scenario they were entering. In the first one I ran back in January, I aimed to create a diagnostic that would show me their argumentation skills, specifically. In that scenario, the students are sitting in a coffee shop. A fellow student sits down nearby. The fellow student notices them using AI and taps them on the shoulder. He wants to know what they think about AI and Writing.

Their job, I told them, was not just to make a statement but to legitimately engage in an argument with their fellow coffee shop patron. The character was designed to take the opposite side of the argument no matter what. If the student says AI is fine for writing, Fictional Coffee-Drinking Student A says it’s terrible. If they say AI is bad for writing, he says it’s fine.

But don’t just make a statement and walk away, I told them. Actually try to convince him. I’d be watching on the back end, and might give them feedback or asked them questions. Show me how good of an arguer you are. And the fun part? He can be convinced. He can also laugh you off.

“So,” I said, “decide on your thesis before you enter the chat. Jot it down. Come up with some supporting ideas. Or don’t. It’s your call. I just want to see how you handle it.”

A few students prepared, reading the dossier closely and asking me a few questions before beginning. Others dove right in after skimming the guide for no more than a minute or two. They leaned back in their chairs, expecting a breeze.

I built in branches below that. The fictional character can be convinced, I decided, by thorough arguments with logical evidence. That seemed obvious. But beyond that, it can also be moved by the user (my students) acknowledging what the fictional character just said, rather than blowing right through it. In other words, they could have the best argument in the world, but if they didn’t actually listen to the person across from them (and show that they were listening), the character wouldn’t change their opinion.

It felt like a good life lesson, baked into an argumentation diagnostic.

On the flip side, if the student does not acknowledge the opposing argument being made, the fictional coffee-drinking student digs his feet in more. He grows more stubborn, even a bit terse. His facial expressions and body language (written in italics in the chat) tell as much of the story as his dialogue does. They also hint to the student which direction the conversation is going. I was interested to see if they would pick up on the furrowed brow and raised eyebrow of the character.

I watched. The students sitting back in their seats began to sit up, then forward, and ultimately hunch themselves over the screen, their fingers pecking the keyboard with a modicum of ire. Then one of them turned to their friend and said, “Yo, this dude is trippin’.”

The students were engaging on a prototype app I had built with a developer friend. It gave me immediate visibility into their performance inside the simulation. So I opened up their dialogues and began reading.

Some students were slowing down, typing in full sentences, acknowledging the counterargument, using anecdotes and examples to support their thinking. I didn’t give them time to do research, so that was the best they could do in terms of evidence. Others were firing off five-word fragments, poor grammar, little-to-no logic or rationale, zero acknowledgment of the counterargument.

Later, we debriefed the experience. Several students found it interesting, engaging, and meaningful. The ones who found it frustrating (several) said, “He was annoying” or “It felt like he wouldn’t listen.” As we kept discussing, the other students chimed in, “But that happens in real life. That’s the point.” Some mentioned feeling like they wanted to win the argument. Perhaps it was an intellectual trigger, I thought to myself.

Still others relayed some version of, “I thought I knew what I wanted to say, but then I realized I didn’t actually know it as well as I thought I did.” They asked if they could start over and do it again. I said sure.

Later, I asked if they wanted to do more, similar simulations throughout the semester. They said yes. I asked how they would feel about having these conversations graded and receiving feedback on their approach. They said “that’s the only thing missing.” I pushed. “Are you sure? This wouldn’t make you feel uncomfortable?” They looked at me like I had two heads. “No, I want to know how I did,” they said.

Ok, then.

Share

Six Months of Experiments

As I worked with my students, I knew we had to try this in other disciplines. So began building simulations for other educators from middle school through graduate school and even into the business world. I partnered with to begin shaping what this approach might look like in a scalable form.

We ran experiments with 12 institutions in total: 7 K-12, 5 Higher Ed. We ran experiments in high school science, math, English language arts, history and social studies, counseling, and language learning. At the Higher Ed level, we covered writing, entrepreneurship, marketing, sociology, and anatomy. They were not all debates, like “The Coffee Shop Argument.” Some took on a different perspective and tested a different set of skills. It often depended on the subject being tested and a host of other smaller variables.

Along the way, we interviewed each educator that ran the experiments and we surveyed the students. We wanted to know what was working and what needed improving.

I am going to begin sharing these stories with you, dear reader, because I believe there is a great deal of potential in this approach. And I’ve continued to build the prototype into a platform that Doan and I plan to launch this August, just in time for back to school. We hope to serve High School and Higher Ed educators looking for a way to redesign assessments and the classroom experience in the age of AI.

We’re hosting a free virtual launch event on Wednesday, August 19th at 1pm EST. We’ll demo the platform and give you a chance to “spar” with one of our Friction Bots, so you can experience what your students would feel. Register here: [AI Friction Labs Faculty Demo Event Registration]

Hope to see you August 19th.

Share

Research Publication Sharing

On a separate note, and I co-authored a peer-reviewed chapter in the forthcoming “GenAI and Higher Education: Ethical Frontiers, Challenge, and Sustainable Pathways“ publication from Elsevier. It comes out July 31st. I have no financial relationship with Elsevier, nor do I make any profit on sales. I haven’t seen the other chapters in the book but will have a chance to read them when I get my copy after July 31st.

Our chapter presents a method for developing AI literacy, along with student data on how the experience changed them. It was a small pilot (21 students, all self-reported) so take the numbers as a first pass, rather than a set of grand conclusions. But we measured three things: skepticism, confidence, and approach to using AI, before and after the activity. We asked – did it make students more or less skeptical of AI? More or less confident? And would they change how they approach the tools going forward? If so, how?

Our thesis was that students who self-report increases in both skepticism and confidence can be considered to have built a higher level of fluency as it pertains to the tools. But skepticism and/or confidence alone doesn’t mean fluent, literate, or aware. For example, an unskeptical and highly confident person could reasonably be considered less literate than a meaningfully skeptical and highly confident person. In the latter case, the confidence is born of the skepticism itself. “I’m confident because I know not to trust.” In the former case, the confidence is born of ignorance. “I’m confident because I don’t believe there is anything to worry about.”

The cherry on top is the approach. Eighteen of the 21 students changed their approach to AI as a result of the activity. The “how” varied. Some said they now think more deeply about what they write into the chat window itself. Some said they trust the outputs less. Cross-reference all three and you have the beginnings of a portrait. An AI-literate person, or a not-literate one.

I’m pretty proud of the work Aimee and I completed. If you do end up picking up the book and have some ideas to share, I’d love to hear what you think. Don’t be a stranger. It gets lonely out here in the desert. Stay tuned. And thanks for being here.

Subscribe now

A Note on the Epigraph

I just finished watching Season One of The Lowdown on FX, a semi-farcical Oklahoman noir from Sterlin Harjo (Reservation Dogs), starring Ethan Hawke. It’s based loosely on the life of Lee Roy Chapman, a Tulsan citizen journalist who famously unearthed connections between one of Tulsa’s founding fathers, W. Tate Brady, and the Tulsa Race Massacre of 1921, among other notable news stories.

The show leans thematically on the work of Jim Thompson, an Oklahoman crime noir author from the 40s and 50s who I was only vaguely familiar with via the commentary of Stephen King [“Big Jim didn’t know the meaning of the word stop. There are three brave lets inherent in the forgoing: he let himself see everything, he let himself write it down, then he let himself publish it.”] Thompson is most well-known for The Killer Inside Me and for reinventing the stylistic norms of noir, even though his work was not widely acclaimed during his lifetime.

I only recently realized how much I actually love noir; I took a class called “Film Noir” in college without having any clue what I was stepping into. Twenty years later, it has dawned on me that its aesthetics are right up my alley. I am also watching Sugar on Apple TV.

In any case, Thompson’s line from the epigraph feels true to life. More than just a commentary on the craft of fiction writing (or a fun, sly comment to make on your way out the door, which I will most definitely be doing going forward), “things are not as they seem” appears to be the plot of every news narrative, stand-up comedy bit, dinner table story, cocktail party anecdote, and speech I have ever consumed. We can’t escape it. If that seems like a grand statement, consider the alternative. Have you ever consumed a story — real or fictional — where things were as they seemed? I’d venture a large bet to say “nay,” because if that were the case, you wouldn’t have a story at all. You’d have a sentence. A pretty boring one, at that.

And that’s how I feel about the AI in Education discussion. Things are not, in any way shape or form, as they seem.

Indeed.

The shift in messaging comes after several years of industry leaders pointing to AI skills as the leading indicator of future employability - and acts as a signal for Higher Ed and K-12 institutions seeking to prepare students for the demands of a fast-changing workforce.

“I can teach the tech skills, I can teach the insurance [domain-specific] skills…what we need is to figure out how to teach the human skills – how to teach future-proof skills that set an employee up for success no matter what domain they find themselves in,” said Steve Jones, Global Head of Data Science and AI at AIG at the AI Summit 2026 hosted by The Urban Assembly and Google.org.

“What we’re seeing is new employees that don’t have the fundamental skills to function in a collaborative team environment. We can filter resumes for experience and degrees, but at that point, we’re looking for people who can collaborate, that demonstrate resilience, that can communicate effectively. That’s becoming a priority for us when we think about the future workforce.”

For the past several years, the message from industry ran in the opposite direction. In May 2025, more than 250 chief executives – including Microsoft’s Satya Nadella, Airbnb’s Brian Chesky, and Uber’s Dara Khosrowshahi – signed an open letter to lawmakers demanding that computer science and AI education be made mandatory for every U.S. student, warning that in the age of AI, children must be prepared “to be AI creators, not just consumers” or risk falling behind.

OpenAI’s Sam Altman has been just as direct, telling one interviewer that the most important specific hard skill a student can learn right now is getting really good at using AI tools – framing it as the present-day equivalent of learning to code when he was in high school. Both LinkedIn and the World Economic Forum, meanwhile, had named AI literacy the single fastest-growing skill workers need to succeed. The directive to educators was unambiguous: teach AI and teach it early.

A Horizontal Workforce

The emphasis on human skills comes at a time when educators are seeking clear directives from industry regarding how to pivot academic programming in the age of AI. Universities have begun adding AI programming at a high rate in an effort to remain competitive in an environment where high school graduates are beginning to question the value of traditional higher education.

But with AI capabilities changing so quickly – and the future so uncertain – industry leaders are increasingly zeroing in on college graduates that can demonstrate “soft” skills that point to success in the workforce.

“We look at skills like negotiation, public speaking, and leadership as the most important in the future,” said Steven Butschi, Director of Education Go-to-Market, North America, at Google on a panel earlier in the day. He pointed to the surge in demand for elevator operators as a representative example, predicting AI will spur a comparable wave of new and unanticipated roles.

Some executives described building a “horizontal” work structure that replaces the traditional vertical structure in corporate environments. Instead of separating employees into columns by their expertise and job function, large employers are increasingly seeking to put employees across sectors into the same room consistently – where the ability to collaborate across fields becomes more important than depth of expertise.

“We can’t just have people doing ‘sensing’ in their own fields anymore,” said Michael Griffiths, US Offering Leader at Deloitte. “We need people to sense, think, collaborate, and design – to be in all those processes at once – and that comes down to group strategy and group decisioning in a way that we haven’t really done before.”

Earlier in the week, industry leaders at the Independent Colleges & Universities of New Jersey Annual Conference hosted by Princeton University articulated similar desires. Executives from major investment banks, pharmaceuticals, and tech companies emphasized the same shift.

“The problem we have is not in finding people with the right degrees, or even the right experience,” said one tech executive on a panel. “It’s in feeling secure that the employee stepping into their new role is going to be resilient, that they’re going to be able to function in a team environment, that they have the durable skills that allow them to navigate a flexible and dynamic working environment.”

Sports and Team Collaboration

Jones added that one of the key indicators his hiring team looks for on an application is whether or not the candidate played team sports in college.

“We’ve had tremendous success in hiring people who have experience in team sports,” said Jones on the panel. “It’s become one of our key differentiators when we have similar candidates in terms of degree and experience.”

An executive at a New Jersey engineering firm communicated the same hiring strategy. Engineers, he noted, have a unique ability to focus on mechanistic details in their work. But without the ability to communicate across teams, those skills drop in value.

“We almost exclusively hire engineers who have played team sports,” he said.

The sentiment has been surfacing in academic conversations for several years. Marc Watkins, who directs the AI Institute for Teachers at the University of Mississippi, has argued that the skills that matter most in an AI-saturated environment are discernment, evaluation, judgment, and reflection – precisely the capacities that “cheap intelligence” can’t supply.

Educators at Elon University, whose free Student Guide to Artificial Intelligence is now used at more than 4,000 institutions across 170 countries, reached a similar conclusion, devoting their 2026 edition to the argument that capacities like critical inquiry, ethical reasoning, and communication are more important than ever, because they’re what let students question AI’s outputs and apply its results with judgment and purpose.

And a chorus of classroom teachers has put it more bluntly: the danger isn’t that students can’t operate the tools, but that they’ll operate them passively. As one California high school teacher told NPR this spring, she feels a responsibility to teach students to interrogate and verify what AI generates – because “if we stop questioning what it says, we can be led to believe anything.”

What Schools Are Building

But university-level responses to AI have often taken the form of new AI-centered academic programming. To be clear, several executives did emphasize that students need guidance around AI as opposed to banning the tool outright, but their commentary on future-ready skills took a different tone.

One New Jersey University at the ICUNJ conference outlined plans to introduce a Professional Studies program in response to the impact AI is having on the workforce. The interdisciplinary program aims to bring students from diverse majors together to analyze the demands of the professional workplace and consider academic opportunities that prepare undergraduates for those experiences.

Northeastern University also has a longstanding College of Professional Studies, built around an experiential-learning model that integrates real-world consulting projects and employer collaboration into the academic program. And multiple university administrators at both conferences indicated that they are developing or expanding “Co-Op” experiences through their Career Centers that move beyond the traditional one-and-done apprenticeship model.

As educators consider the impact of AI on the traditional educational experience, the shift in messaging from industry leaders is likely to serve as a key signal for the development of new academic programming that may or may not be AI-centric – but will likely be human-centric.

I wrote a piece for The Hechinger Report this week arguing something I’ve believed since ChatGPT first shipped but have been reluctant to say this plainly: AI is a drug. Not as a figure of speech for how it makes you feel — I mean it the way we talk about anything freely available, mildly addictive, and being consumed in enormous quantities by people who’ve had no education about what it does to them.

The reflex with drugs is to ban them, and the reflex in a lot of schools right now is to ban AI. But we already ran that experiment with “Just Say No,” and we know how it went. What actually built resistance was education — honest information about effects, dosage, and risk.

Which raises the obvious question for anyone teaching right now: if not abstinence, then what’s the right dose? I have a partial answer, and it came from a classroom moment I didn’t expect — I went in braced for the worst thing AI could do to my students, and watched something close to the opposite happen instead.

That’s the piece. Read it here, and if it’s useful, send it to someone who’s still deciding whether to ban the thing.

TEACHER VOICE: AI is an addictive drug that must be researched, studied and confined

I wrote this piece in May 2024.

This week, Pope Leo XIV released his first papal encyclical — a letter meant to define his entire papacy — and built it around the same question I was asking.

The Tower of Babel, if you don’t remember the story, is about humans overreaching — building something so powerful and prideful that it collapses under its own ambition. The Pope used it to describe exactly what AI risks becoming: “a top-down, grandiose project where decisions are driven by pride, profit and a push for homogenization.”

He put it simply: humanity’s choice is between “constructing Babel or rebuilding Jerusalem.”

I’ve been thinking about that choice for two years. Here’s what I wrote back then.

AI and The Tower of Babel

I work at a Catholic school. Though I was raised Catholic and believe in God, I don't strictly follow a particular faith and would probably even fall into the category of “lapsed.” Over time, I have found valuable lessons in various religions and developed my spirituality, like many, through diverse experiences. However, my Catholic upbringing remains central, shaping my morals and ethics.

One story from my upbringing came to mind yesterday when OpenAI released ChatGPT4o and all its new features, especially the language translation. By now you’ve likely seen the video, but if not, here it is. The voice translation feature could change how we communicate across languages, races, and cultures.

This made me think of the Biblical parable of the Tower of Babel. I'm not a theologian, but I asked a Catholic monk to check this story. Here's how I understand it:

In Genesis, the Babylonians wanted to build a great city with a tower that reached the heavens. God disrupted their plans by confusing their language, so they couldn't understand each other. The city was never completed, and the people were scattered across the earth.

This story is meant to explain why we have different languages. It's also why we say someone "babbles" when they don't make sense. God mixed up our languages to humble us, reminding us that we can't reach heaven by our own efforts.

This parable can be interpreted in various ways. For our discussion, the most relevant analysis examines the appropriateness of humans striving beyond their limits. It also explores the tension between human ambition and respect for divine authority.

Now, think about this new AI technology. What happens when there are no language barriers? Many people believe this is positive—imagine what we could achieve if we could all communicate without problems, they say. Perhaps it could even end conflicts between countries…

But isn't this like trying to build a Tower to Heaven again?

I've mostly avoided the deep, philosophical discussions about AI because it's beyond my control. AI will be a significant part of society, whether I like it or not. Instead of debating if AI is good or bad, I focus on how we react to it. This mindset, inspired by stoic acceptance and the urging of Geoffrey Hinton, helps me cope.

However, the recent release of ChatGPT-4 has me thinking deeply again. Ethan Mollick says you need "three sleepless nights" to fully grasp AI's impact. I'm already at four, and I don't see it stopping.

How many sleepless nights will GPT-5 bring?

Frustration and Humility in the Learning Process

In the summer of 2019, I traveled to the Tibetan village of Ritoma in Western China to coach a semi-professional basketball team. This team was composed of nomadic Tibetan men employed by the high-end clothing company Norlha Atelier. It was one of the most amazing experiences of my life.

For a month, I practiced with Dugya Bum and Chuchon Jyap, the two full-time players on the team. During the offseason, they worked for the clothing company and no longer lived as nomads. But for the first week, we traveled with five other "hired guns" to the city of Luqu in Western China. There, they played in a local cash tournament that was broadcast online across the country.

I attempted to coach them. This story deserves its own blog – or even a book – but it resurfaced in my mind last night while laying in bed and watching ChatGPT4o effortlessly translate English into Italian on the fly.

I couldn’t help but remember how intensely I struggled that first week in China while in Luqu for the tournament. Of the seven players I was supposed to coach, only one spoke a bit of English. I had planned to use cell phone translation apps to cross the language divide where necessary, but that plan was dismantled when I found out that six of the players spoke only Tibetan – not Chinese – which did not exist as an option on any of the translation apps I had downloaded ahead of the trip.

Fortunately, one player spoke Mandarin. He became my main point of contact for that week. We handed my phone back and forth to communicate. For the rest of the time, I relied on hand motions, facial expressions, and speaking English to Dugya Bum, who could understand me but only spoke in broken English.

This experience was incredibly humbling. I had left JFK airport in New York City expecting to arrive in Tibet as a paragon of basketball skill and knowledge. I thought the players would be grateful for my volunteer time, and that I would lead them to victory in the tournament.

Instead, we fumbled through interactions and never made it out of the group stage. Later, the proprietor of the team – the owner of Norlha Atelier – asked me, "So...what happened?" in confusion. The team usually placed in the top three of each tournament, bringing home a cash prize that supported the team's continued existence and growth.

The humbling experience was not just due to my inability to communicate but also my pride. I was so convinced of my own value -- and so unaware of my flaws -- that I missed the glaring gaps in my approach. I spent so much time frustrated by my inability to make my experience match my the picture I had created in my imagination that I missed the beauty and joy of the reality in front of me.

Pride does come before the fall. It was only after enduring prolonged periods of solitude—between our games—and experiencing a condensed version of "the dark night of the soul" that I found a renewed sense of purpose to make a meaningful contribution. If my dreams weren't going to be realized, perhaps I could carve out a role for myself in this challenging reality where I could be an additive force.

It wasn’t easy, but I did find that joy. I learned to step back where I wasn’t needed, took opportunities to help where possible, and ultimately found a way to bond with the team through the confusion of the language barrier.

Everything about this experience exhausted me, and that’s what made it so great. I wonder, with AI making communication easier, what will happen to this kind of discomfort? Is it better for things to be easy?

Subscribe now

The Disappearance of Languages

Watch this video again.

Is this a good thing? If future technology lets us wear earbuds that translate languages instantly, will we still need to learn new languages? I remember when Spanish was considered "the language of the future." My grandmother, a French teacher for thirty years, spent a week living in Mexico with a host family as an adult in the 1980’s to become fluent in Spanish for that exact reason.

Now, what is the language of the future? Is language itself becoming less important? If I want my child to be bilingual, could I use an earbud to translate as I speak English, helping them learn Chinese at the same time? What happens when my teenager curses at me in Mandarin ten years later, and I have no idea what they just said?

Many educators believe that some struggle is essential in learning. Thoughtful blogs about the need for “friction” in the learning process have popped up here, here, and here (and more). This is a good thing and should continue; educators need the language to explain why critical thinking and analysis are vital. They need a lens through which to explain the need for the conservation of friction in the learning process.

The key to preserving challenge and resistance in the classroom lies in the interactions with AI themselves. We need to focus on teaching each student to engage critically with AI and making thoughtful interaction part of the learning process — rather than working around AI on the whole. By monitoring, evaluating, and grading these interactions, we can keep the necessary learning and critical thinking challenges in place to help our kids and society progress. Otherwise, we risk becoming mindless drones asking AI to do everything for us, perhaps even building a faulty tower to the sky once again, only to be struck down by a power vaster than anything we – or AI – could imagine.

Epilogue (May 2026)

When I wrote that last line two years ago — “building a faulty tower to the sky” — I didn’t have an answer. All I had was a fear — one that has driven my work ever since.

Pope Leo’s encyclical gave me the other half of the frame. The choice, he wrote, is between constructing Babel and rebuilding Jerusalem. Babel is top-down, pride-driven, homogenizing. Jerusalem is diverse people working together, preserving what matters, building something humans can actually live in.

That’s the vision I’ve been trying to build toward. Over the last two years I left the classroom and started a company that builds AI simulations for educators — tools that put students into difficult, ambiguous situations and make them think their way out. Not removing friction. Preserving it. Not relying on AI for homogenized information. Creating fictional environments where a human being can grow.

Soft skills. Fictional environments. A new foundation for education and the construction of a human being.

The struggle is the point. That’s Jerusalem.

The Pope just named the stakes. I hope educators are listening.

Discussion 1: Whether and when to use AI — an ethical question.

Discussion 2: How to teach students to use AI better, so they protect their own cognition and creativity — a question of maturity and skill.

These two conversations talk past each other and produce stasis. Today I want to name a second conflation that creates the same problem.

Here’s the situation on most campuses right now. A new College Board survey of more than 3,000 U.S. college faculty (February 2026) found that 74% of faculty report students using AI to write essays or papers, and 45% hold an overall negative view of AI in higher education. A separate survey from the American Association of Colleges & Universities and Elon University’s Imagining the Digital Future Center put an even starker number on the table: 95% of faculty believe AI will make students dangerously over-reliant on technology, and 78% say cheating on their campuses has increased since AI became widely available.

Faculty are scared, frustrated, and increasingly burnt out by a problem that keeps getting bigger.

But here’s the thing. The conversation meant to address that problem keeps splitting in two — and neither half can move until someone separates them.

Two Problems. One Meeting. No Progress.

Teaching AI Literacy and Preserving Assessment Integrity are not the same problem.

Assessment integrity has deteriorated because of AI. Students can more easily cut corners on assignments, which destroys an educator’s ability to evaluate student thinking in any digital space. This is a real problem with real institutional stakes.

AI Literacy is a different project. It focuses not on preserving assessment integrity but on teaching students to be critical thinkers in the chat — to reflect, push back, and use AI in a way that builds their own cognition rather than replacing it. Researchers like Ethan Mollick at Wharton — author of Co-Intelligence and one of TIME’s Most Influential People in AI — have argued consistently that how students engage with AI determines whether the technology strengthens or atrophies their thinking. Marc Watkins, writing at Rhetorica, makes a parallel case: pedagogical approaches need to blend AI engagement with analog methods that protect human agency and judgment.

Both problems deserve serious attention. But they are not the same problem. And when they sit in the same meeting, neither gets solved.

What Happens When You Try to Solve Both at Once

Some AI Literacy practitioners — myself included — have tried to put the two together. What if we taught AI literacy in the context of our assessments? Might that preserve assessment integrity while building the literacy we need? Two birds, one stone?

It is possible. My framework for comparative transcript analysis has succeeded in both changing student perception of AI use and providing educators with a viable window into student thinking in context. You can read more about its first application in the Foreword of the Field Guide to Effective AI Use.

But as I’ve sought to scale the mechanism, the conclusion has become clear: trying to accomplish both at the same time puts a massive cognitive burden on both the educator and the student. In Aimee Skidmore’s and my Elsevier chapter (due to publish later this year), we document what this looks like in practice — four weeks of sustained effort to run one coherent cycle. It works. But it’s not scalable, and it’s not realistic to ask of most faculty.

The EDUCAUSE 2025 AI Landscape Study confirms the structural problem: only 22% of institutions have an institution-wide AI strategy, and for 55% of campuses, AI policy is happening in fragmented pockets. The AAUP reported in 2025 that 71% of faculty say administrators overwhelmingly lead AI conversations with little meaningful faculty input — and only 20% of colleges and universities have published any formal AI policy at all.

The institutions that are spinning their wheels aren’t failing because their people are bad. They’re failing because they’re asking one committee to solve two fundamentally different problems at the same time.

The Fix: Split the Problem

Here is what the separation looks like in practice.

Two Committees

Instead of one AI Committee or Working Group, higher ed and K-12 institutions should create two:

• An AI Literacy Working Group focused on what mature, critical AI use looks like — how students engage in the chat, how to build metacognitive habits, how to teach AI in its broader societal context.

• An Assessment Integrity Working Group focused on how to evaluate student thinking in an AI-saturated environment — portfolios, process-based assessment, project-based learning, conversation-as-artifact. These solutions do not need to involve AI at all.

The key insight: assessment integrity can be preserved or rebuilt without involving AI literacy at all. AI literacy can be built without navigating around AI cheating. These are separate tracks aiming at separate outcomes.

Two Sets of Experiments

For the AI Literacy Working Group: experiments that center on what it means to use AI in a way that builds the user. Teaching skills in the chat. Teaching students to place AI in its broader societal, economic, and political context. Training metacognition, not just prompting.

For the Assessment Integrity Working Group: experiments focused on portfolios, project-based learning, and process-based evaluation mechanisms. The EDUCAUSE ALTL framework (October 2024) and work from researchers like Laura Kristen Allen and Panayiota Kendeou at the ED-AI Lit Project provide strong scaffolding for the literacy side. Academic integrity scholars Tricia Bertram Gallant and David Rettinger address the integrity side directly in their 2025 book The Opposite of Cheating — the thesis being that academic integrity is fundamentally a teaching and learning issue, partially solvable through good pedagogy.

Think about what we’ve been doing instead. We’ve put two groups of educators — with genuinely different goals — in the same room, pointed them at the same deadline, and then wondered why progress is so slow.

Why This Matters Now

The frustration simmering across campuses is real. Research published in the Journal of Academic Ethics argues that AI didn’t create the assessment integrity crisis — it exposed and accelerated a structural vulnerability that was already there: massified, standardized assessments designed for administrative convenience rather than pedagogical fidelity. AI made that vulnerability impossible to ignore.

The answer isn’t to bolt AI literacy onto the integrity crisis or vice versa. The answer is to give each problem the focused attention it deserves — separate working groups, separate experiments, separate success metrics.

I’m not under the illusion that this idea travels quickly. For faculty still in the anger, denial, or bargaining stages of the AI transition, it may not land yet. But for the educators already out front — already asking the harder questions — this distinction is useful. We’ve been splitting our focus in two and aiming it at a single goal, when the two tracks are actually trying to achieve something different.

The call to action? Name it. Separate it. Make progress.

A note on how I’ve organized my own work along these lines: AI Literacy Partners (litpartners.ai) is my consultancy — that’s where I do AI literacy work with institutions. AI Friction Labs (aifrictionlabs.com) is a separate product built specifically to address the assessment integrity problem. Two different projects, two different websites, for exactly the reason this post argues. If the assessment integrity track is the one you want to follow closely, the AI Friction Labs newsletter is the place to do that — sign up at aiFL Newsletter Sign Up.

The em-dash. The antithetical. Now, the rule of threes.

Each of these rhetorical staples has experienced a certain “AI death knell” over the last three years. We notice it first, then talk about it on a cultural level, then it gets discarded.

First, arguably, it was the em-dash. Writers across the social and cultural universe have expounded on the immense frustration rising from the fact the use of an em-dash is almost automatically tabbed as “AI writing,” even when the writer didn’t use AI. They edit their own writing with this in mind, regardless of whether or not the em-dash itself was effectively placed.

I’ve told this story before – but I have always been a big fan of em-dashes. My father, a lawyer and the person in my life who taught me the most about effective communication, uses them ad nauseam. Through the process of osmosis, they became a staple of my own writing. And when I was a cub reporter at Thomson Reuters in 2012, I nearly came to blows with my editor over their use.

“I hate those dashes,” she said, commenting on one of my weekly columns. “It’s poor writing.”

“But a comma doesn’t work there,” I said. “Neither does a semi-colon. And a period is too abrupt. It’s part of the same thought – not a new sentence.”

I won that argument, if only because I was too stubborn to learn a new way to communicate.

And yet here we are, editing them out anyway – not because of a stubborn editor, but because of a machine. From my perspective, that can only be described as a damn shame. To think that an effective grammatical device has been culturally appropriated by a device with no true understanding of what it is doing is tantamount to watching your own personality be mimicked by an inanimate object. And the idea that an inanimate object could have such a profound effect on society that it forces me to change my personality is, frankly, a bitter pill I’m not ready to swallow.

But developers are watching. Because notably, each time that our awareness of a new AI-ism surfaces in the zeitgeist, AI companies quietly shift their own models away from those habits – or modify them in some unrecognizable way. The AI-produced em-dash, for example, has gotten longer in recent months. What used to be a “–” is now a “—.”

The fact that AI is now trying to sound less like AI feels more like a dystopian irony better suited as a world-building mechanism in a science fiction novel than a reality that we as human beings have to navigate around.

But here we are. And new tendencies are surfacing. AI is now coming for the rule of threes – that well-worn strategy of delivering an argument in staggered, sequential packs of three sentences, phrases, or even individual words.

The implications of AI’s shifting communication style are manifold. Below is an attempt to capture the cultural impact of this trend, for better or for worse.

Share

Alien Language

University researchers – in conjunction with Google DeepMind – recently mapped the linguistic style of AI models in an attempt to capture its potential impact on the way we communicate, absorb, and digest information across the digital landscape.

They found that AI language lives somewhere outside of any common human linguistic approach. It’s an alien language, in other words.

Think about that: An AI is trained on the way we speak and write (primarily write.) It tries to sound like the version of us that will maximize positive responses. In doing so, it creates an entirely new form of communication that is not necessarily ineffective, but is entirely foreign.

“Human revisions exhibit smaller and more diverse shifts, whereas LLM revisions produce large semantic shifts strongly aligned in a common direction, and to a region of space not previously occupied by any human-written essay,” they wrote.

Comparisons to Frankenstein have littered the cultural landscape since the dawn of commercial GenAI in November 2022. The UW/Deepmind study would seem to indicate that AI developers have succeeded in creating a Linguistic Frankenstein.

If we are using Shelley’s novel as a prism, we have to consider where the fictional monster ended up. Dr. Frankenstein sought madly to contain his creation, but in so doing – as Anne Mellor notes in “Mary Shelley: Her Life, Her Fiction, Her Monsters” – he became indistinguishable from it.

How might that manifest itself in our modern day predicament?

We don’t need to look to the future to find out. High school and college students, in particular, report altering their own writing style in an attempt to not get flagged by AI detectors. They are, in sequence, writing their own essay by hand, running it through an AI detector, and then using the detector’s report to, potentially, “dumb down” their own writing output to avoid academic integrity violations.

This is not a near-tragedy. This is an actual tragedy.

As a writer myself, I have increasingly turned to AI to assist with longer-form, more complex attempts at communication. What started with emails in the fall of 2024 turned into assistance with blogs and even some of my graduate-level work (with approval, of course) by the fall of 2025.

But everybody has their limits. For me, the incessant reliance of AI models on the rule of threes feels like the straw that broke the camel’s back.

Share

Larry, Moe, and Curly

If/when I use AI for writing now, I always deliver a list of “Don’ts”:

1. Avoid antitheticals

2. Use minimal em-dashes (I just can’t quit them)

3. Avoid short, punchy sentences in lists of three - or -

4. Avoid the rule of threes altogether

5. Avoid “Here’s the thing” or “Here’s why that matters.”

6. Avoid adverbs

(That last one requires a post of its own. Claude’s tendency to tell me that my idea is “genuinely fascinating” has become enough to trigger a new level of fury. It’s enough to abandon the platform. But one cannot deny the psychological impact of these effectively placed adverbial constructions on the mind. They are, I believe, an underrated and under-discussed factor in creating dependence, addiction, and even psychosis in human users. Put simply, it feels good to hear someone say that your work is “profoundly innovative.”)

But even the most closely guarded AI can’t break its bad habits. And when I read a piece of AI writing – whether in my own interface or on someone else’s LinkedIn – I can’t deny that the rule of threes is effective. It’s like a salve for the brain.

Alvin, Simon, and Theodore. Snap, Crackle, Pop. Stop, Drop, and Roll.

These are sticky because past writers discovered they were sticky. They tried out new things, new modes of communication, new strategies. And when they worked, they kept using them.

Which leads me to my final conclusion.

Share

We’re Gonna Find A New Way To Write

The physics of displacement indicates that when one object is moved by another object, it finds a new place to go. Our communication styles, it would then seem, must find new territory to inhabit.

To truly sound different than AI, a writer – or speaker, for that matter – has to innovate. It’s not enough to avoid em-dashes, antitheticals, and the rule of threes. This post is littered with them, and yet Claude was provided zero access to this piece. Innovation in this space, then, would need to center around three questions:

1. Have I been effective?

2. Do I sound like AI?

And most importantly,

3. Can you hear me?

This last question is the most important because the quality that AI most lacks is “voice.” Voice is a quality that is developed and sits so deep within a writer’s consciousness that it often cannot be identified, defined, or even copied. It is uniquely you – and, luckily for us, human beings contain so many multitudes that not even a superpowered language model can capture them all.

I posit that the development of voice – whether in writing or verbal communication – will become a staggeringly important human quality in the future. It will happen both naturally and intentionally. Some will notice it subconsciously. Others will point directly at it.

Either way, all of this points to one of two developments in the field of communication:

1. A more distinct separation of writers/communicators that “have voice” and those that don’t. I refer to this as Luxury Language. Much like cars, handbags, and clothes – a market for Luxury Language will surface in response to the AI Language Assembly Line. This is a cultural manifestation of the physics of displacement.

2. The creation of an entirely new method for communicating the self.

This second development is the one I am most interested in. In order to come to fruition, such a new method would require a new medium unto itself. It can’t exist within our traditional mediums, because AI is studying our traditional mediums. It will have to live on its own – in a quiet, unseen pocket of human life that escapes the Sauron-like gaze of AI models.

This is the moment of the piece where I am supposed to tell you what that medium is. But I am not going to do that. Why? It can only be one of three reasons:

Either:

1. I don’t know at all/I’m clueless

2. I know, but I’m scared to say it (because I don’t want to be publicly wrong)

3. I know, but I don’t want AI to know

Take your pick. Any pick. Because the truth of the matter (in a post-truth world) is that the truth doesn’t really matter.

All that does matter – is how you say it.

But a growing number of educators have cleared that hurdle and are standing in the room that comes after. The light is yellow or green. A teacher has decided that AI use is going to happen in their classroom, under some set of conditions, for some defined purpose. And now they are looking at a blank page where the guidance should be.

The instinct in that room is to write an ethics policy.

It is the wrong instinct, and understanding why is the difference between a policy that does something and one that produces a document.

Nearly every educator I speak with who has reached the yellow or green stage asks the same set of questions: how do I define ethical AI use for my students? How do I teach it? How do I evaluate it? The questions are reasonable. The framework underneath them needs work.

Ethics is the right lens for the stoplight. Whether to allow AI at all, and under what circumstances, involves questions that belong to the community: the environmental cost of large language models, intellectual property and the integrity of other people’s work, the institutional standards a school sets for itself. These are communal decisions. Communal standards are the right tool for making them.

But once a student is sitting in front of an AI and engaging with it — once the light has turned and the decision has been made — the question of how that engagement unfolds is not an ethical question. It is a maturity question. Applying a communal ethical standard to the quality of an individual student’s cognitive behavior is an error of mis-categorization, and it is why most AI use policies, however carefully written, are not doing what their authors think they are doing.

Ethics Are Communal

Ethical frameworks are meant to govern behavior across a community. Every application of ethics in societal life, though, runs through individual circumstance.

Take violence. By most standards, it is wrong to commit murder, wrong to commit acts of violence. And yet the standards of fairness we maintain are predicated on self-defense, psychological wellness, and the specific facts of each situation. When a person commits a violent act and claims self-defense, a trial ensues. Circumstantial evidence is presented. What was going on before the act? Did the situation require violence to survive? Who was the aggressor, and to what degree? Was the defendant capable of knowing what they did was wrong?

Each of these questions takes into account the specific characteristics of the situation and the individual. The lawyers tell a story. The judge considers the person and the event in context. The jury determines whether the specific incident met one standard or another.

When the UnitedHealthcare CEO was shot in December 2024 and a significant number of people cheered. As abhorrent as it was, it points to something important that we seem to be forgetting about ethics in and of themselves.

Thousands of years of legal and philosophical development, and we still cannot agree on whether or not murder is always wrong. Why? Because we automatically take circumstances into account, and our evaluation diverges. People thought the murder was “right.” To be clear, I am not in their camp. But they thought the murderer was justified based on how health insurers have treated customers over the past few decades.

As a broad conclusion, these examples point to the fact that ethics provide a set of communal standards, but they are always applied via individual circumstance. The problem with Ethical AI Use discussions, it seems to me, is a failure to recognize and apply similar circumstantial evaluation practices when it comes to determining whether a student meets the “Ethical AI Use” standard.

How does this problem manifest itself in an AI engagement? Well, for one, AI use is highly circumstantial as well. Every AI chat is different. The record of actions, choices, and responses is individual in nature, and the circumstances that led a student to engage the way they did are unique and specific.

Was this a 9th grader? A freshman in college? A law student? A 5th grader? What were they using it for — brainstorming? Research? Drafting? How much instruction had they received beforehand? What scaffolds were they given? How much time? What were the incentives?

A single educator will see these variables and respond; “But my students all fall within the same set of circumstances?” Ah, but do they? Let’s go further.

Assume the answers to all of those questions are identical for a class of 25 fictional 9th graders. Same situational variables, every student.

What is Student A’s cognitive level compared to Student B’s? How much has Student C used AI compared to Student D? What is Student E’s reading level compared to Student F? Student G’s writing level compared to Student H? How much feedback on AI use has Student I received compared to Student J? How many opportunities has Student K had to practice meeting the standard compared to Student L?

That list does not begin to cover it. It doesn’t even account for the differentials in the AI’s responses themselves. Now imagine applying an ethical standard to all 25 students at once, without taking their individual circumstances into account. This is what an AI ethics policy asks teachers to do. And this is why it fails.

Subscribe now

The Essay and The Chat

Teachers already know how to do what I am describing. They do it every time they grade an essay.

Consider how the average teacher evaluates a student’s ability to meet a writing standard. Science teachers: read this as if I am talking about lab reports.

Jimmy and Jenny are in the same class at different performance levels. Jimmy struggles with reading, Jenny reads above grade level. Jimmy’s organization is behind the curve, Jenny has received support at home that keeps her diligent. When I grade Jimmy’s paper I apply my rubric, find the gaps and the highlights, mark them and record the points.

Then I take into account Jimmy the person:

Based on where he was before this paper, did he improve? Did he try? Is there evidence of effort? Did he take my feedback and apply it in some fashion?

Those last questions are maturity questions.

In academic terms, maturity equates to effort. I cannot expect Jimmy to ace the essay, but I can expect him to try. If he does, he deserves credit. He may have fallen short of the standard, but that does not make him a bad person. Nor does it mean he did not grow, or try, in the context of the assignment. He showed maturity, and for that he deserves credit.

Now imagine evaluating that same paper through the lens of ethics. Imagine saying to Jimmy: you wrote this paper, but you did not meet my standard. Therefore, you are unethical.

That sentence sounds absurd when applied to a writing assignment. It sounds equally absurd when applied to an AI chat — and yet that is the framework most schools are currently building and seeking to apply.

What the Maturity Standard Actually Looks Like

The above graphic visualizes what it might look like to design evaluation standards for a student’s AI use. You certainly can - and probably should – have a standard set of criteria for the whole class if you are assigning a “yellow” or “green” assignment.

But if you are subsequently reading their AI use and asking yourself whether they met your standard, it isn’t enough to execute a 1:1 comparison.

Instead, your evaluation criteria should be combined with these (and other) highly individualized set of evaluation mechanisms. From here, it becomes much easier to say to a student something like the following:

“I like how you did X. That very clearly matched the criteria I laid out for this AI engagement. However, at one point in the dialogue (Y), your approach appeared to veer away from what we have defined as meaningful engagement for this AI use. In the future, try Z.”

Sound familiar? It’s the Glow-and-Grow framework, applied to an AI chat. Complicated? I dare say not, especially if you have already determined your criteria for that particular engagement (research, brainstorming, drafting, sparring — all taking into account the level of scaffolding provided).

This is how I provide AI use feedback to students in my Rhetoric and Inquiry class at Fairleigh Dickinson University — and it works. Students understand what I want them to do, and apply it the next time around.

Additionally, please notice that nowhere in this feedback did I imply that the failure to meet the criteria indicated an ethical failing. I wouldn’t tell Jimmy (or Jenny) that they were a bad person if they failed to meet my essay requirements. So why am I telling a student that they committed a moral failing when they struggle to meet my metacognitive engagement standards with AI?

Share

AI Discussions at Schools

One of the reasons schools stall in the policymaking process is that their committees are collapsing two distinct conversations into one. Educators and administrators keep asking what it means to use AI ethically, when the more useful question — once you are past the stoplight — is what it means to use AI maturely.

The stoplight itself is an ethical decision. A red assignment means the educator has determined, for whatever reason, that AI use is inappropriate in this context. That determination belongs to ethics: it is a communal judgment about what is acceptable.

Once the light is yellow or green, the ethical question has been answered. What remains is the maturity question: given that a student is engaging with AI, given everything we know about where they are developmentally, did they engage at the level they are capable of?

Institutions that recognize the distinction can structure their conversations accordingly. The stoplight debate — whether and when — belongs in one room, governed by communal ethical standards. The guidance on how to teach and evaluate AI use belongs in another room, governed by the maturity framework. Both conversations need to happen. Running them simultaneously is why so many committees produce documents that satisfy no one and guide no one.

The Conditions That Make It Possible

The maturity frame has direct implications for how you design AI interactions in the first place. The design of the interaction matters as much as the student’s response to it. When students encounter a tool that removes all resistance, they find the minimum viable input and stop. That is a rational response to the incentive structure. When they encounter an interaction that maintains resistance — one that requires thinking, adapting, and persisting — the transcript reveals something no essay can: how they engage with difficulty. And that, in turn, reveals maturity.

AI Friction Labs designs interactions that meet these conditions. They are structured in a way that makes meaningful engagement much clearer to the student. They reduce the cognitive burden on both the educator and the student as it pertains to defining the “how.” In turn, they develop the “soft” or “future-ready” skills in context. You can see how they engage with simulated real-world difficulty and provide an arena for them to engage with AI maturely.

Reach out at https://aifrictionlabs.com to learn more.

Ethical AI Use is about “Whether and When.” Mature AI Use is about “How.”

When you have your next AI Committee meeting, split the two. You will find that the conversation gets shorter, the guidance gets clearer, and the teacher standing in front of a student’s chat transcript finally has a framework for understanding what they are looking at.

Weizenbaum spent the next forty years warning people about what he saw that afternoon. He called it the illusion of understanding. Since then, every few years, another AI researcher published a paper saying the quiet part out loud: these systems are not thinking. They are performing. When you chat with one, you are not interacting with a mind, but walking onto a stage.

We keep treating those warnings as reasons to back away. I think they’re instructions for how to walk in.

The diagnosis has been made

The modern version of Weizenbaum’s warning starts in 2021 with Emily Bender, Timnit Gebru, and colleagues publishing “On the Dangers of Stochastic Parrots,” — arguing that large language models stitch together probabilistic sequences of words without reference to meaning. The paper cost Gebru her job at Google. Years later, Sam Altman’s reply was a tweet: “i am a stochastic parrot, and so r u.”

In 2023, Murray Shanahan and colleagues published “Role Play with Large Language Models” in Nature. Their thesis is that dialogue agents aren’t actors reading a script, but are improvisational performers staying in character. The chat, then, in essence, is role-play through and through.

Shannon Vallor’s The AI Mirror, out of Oxford last year, takes the metaphor where it wants to go. These systems aren’t minds, she argues. They’re reflections of our collective intelligence, shallow in the way a reflection is shallow. Not deceptive. Just flat.

Sixty years. Multiple fields. Peer-reviewed consensus. When you open a chat window, you’re not interacting with a mind. You’re walking onto a stage.

The warning is defensive by default

Every one of those sources presents the simulation insight as a problem to be guarded against. Don’t project. Don’t be fooled. Don’t anthropomorphize. Keep your distance from the mirror.

That instruction is reasonable, if not wise. But it’s also limited because It treats the simulation as a trap — something to be recognized so you can avoid it. But if every AI interaction is a simulation, there is no version where you avoid the stage. You cannot opt out of walking on. The only real question is whether you walk on deliberately or by default.

Most people walk on by default. And the default has a shape.

The scene that ships in the box

ChatGPT ships with one scene pre-loaded: the Helpful Assistant. Eager, obliging, confident, fast. Most users never leave it because they don’t know they’re in it. The scene is invisible until somebody names it.

The Helpful Assistant is fine for tasks. It is the worst possible scene for thinking. It rewards passivity, validates half-formed ideas, and produces polished output that removes any need for you to have done the cognitive work yourself. You ask a vague question, it gives you a confident answer. You share a half-formed idea, it expands it. You float a thesis, it supports it. Nowhere in that exchange did you have to do the work of specifying, defending, or revising what you actually think.

Leticia Britos Cavagnaro at the Stanford d.school has been making a version of this point for a while. Her framing is “what happens when we turn the tables on AI and use it to develop our human intelligence?” The question presumes the default orientation is the wrong way around. I think it is.

Choose the scene

The conclusion here is that the mirror is very real. The role-play is real. The simulation is real. Stop arguing with the researchers, because they’re right. Instead, ask a powerful question: which simulation do I want to be in?

An editor who kills your paragraphs. An investor who refuses to buy your pitch. A skeptical expert in a field you think you understand. A client who will not accept your jargon. A devil’s advocate who takes the position you cannot stand. A mentor who is tough because the work needs it to be.

Each of those is a different scene. Each puts you in a different role. Each one asks something different of you. None of them produces polished output. All of them produce a better version of your thinking than the Helpful Assistant ever will.

The point is not to convince yourself the AI is real. You know it’s not. The point is to cast the mirror in a scene that demands something of you. The system cannot think for you. But the scene you put it in determines whether you think at all.

Lean in. Pick the harder room.

What follows from this

Once you see AI chats as scenes you choose, a few things shift at once. Prompting stops being a trick and starts being casting. Evaluating AI output stops being about accuracy and becomes about whether the scene did its job. That’s an expertise question, not a question of factual accuracy. Assessment — the thing schools and companies are currently panicking about - stops being about what the AI produced and starts being about what the user did inside the scene.

That’s what my partner Doan Winkel and I are working on over at AI Friction Labs. Story-based, challenging AI simulations that put students in the arena where they can practice, train, and get feedback from you - the expert – in a safe environment.

Landing

ELIZA was not a mind. ChatGPT is not one either. Every time you open a chat, you’re walking into a room with a mirror. The mirror will reflect whatever you bring into the room.

So bring something harder than “please help.”

More at aifrictionlabs.com

If you’re a teacher, you already know the essay assignment is broken. Your students didn’t stop caring about writing, but now they have a tool that writes for them, is faster than they are, is pretty good, and you can’t prove they used it. Detection tools don’t work and turn you into a cop in your own classroom. Oral exams don’t scale to a class of thirty. “Integrate AI thoughtfully” is advice that sounds reasonable until you try to do it without losing the one thing you were actually trying to assess: whether your students can think.

The problem teachers face isn’t that AI exists. The problem is what it’s optimized for. ChatGPT, Gemini, Claude, these tools are built to be helpful, which means they resolve ambiguity, supply missing information, and sand down every rough edge in a student’s thinking before it ever has a chance to become something real. That’s a perfectly sensible design for a productivity tool. For learning, it’s a catastrophe, because the research on productive struggle is unambiguous: the difficulty is the learning.

When AI absorbs the cognitive load, the student doesn’t build the skill. They produce an output. Those are not the same thing, and the gap between them is exactly what’s eroding in classrooms right now.

I’ve spent the past two years writing about this problem from every angle I could find. What it looks like in assessment, what it does to faculty who are trying to respond to it, what happens cognitively when the friction disappears from a learning interaction, what we actually mean when we say we want students to think critically. Each of those posts was a version of the same underlying question: if our traditional assessments and learning tools are no longer a reliable window into student thinking, what replaces it? But the thoughts and concepts I was offering were slowly morphing into a solution, or an answer, perhaps.

That answer is AI Friction Labs.

What the platform does

Friction Bots are character and scene-based AI interactions designed around specific learning outcomes, built to put skills in context rather than in the abstract. A student in a finance course defends a deal decision to an investment committee member who keeps pressing on the assumptions. A marketing class tries to persuade a potential buyer who has no particular interest in switching products and says so. A public speaking student presents a project idea to a town hall and then engages with a community member who raises objections that don’t resolve easily. In each case, the simulation is built with a dossier, a set of resistance behaviors, and a defined scenario, and the bot maintains that resistance across the full interaction rather than softening once the student produces something plausible.

The transcript of the interaction is the assessment. Educators get full visibility into how the student moved through the scenario: where they adapted their approach, where they got stuck, where they recovered, how they handled being challenged on something they thought they understood. AI-assisted grading surfaces patterns across a whole class and generates feedback that would take hours to write manually. The tool works as a low-stakes practice environment, as a formative assessment, as a way to introduce a skill in a real context before students encounter it in the world, or as a formal summative evaluation. What it cannot be used as is a shortcut, because the interaction is designed so that there isn’t one.

This matters beyond the classroom, too. The skills that AI is eroding are precisely the skills that are becoming more valuable as AI takes over the information layer of knowledge work. The ability to hold a position under pressure, adapt an argument when it isn’t landing, communicate clearly with someone who is resistant, regulate your own emotional response when a conversation gets difficult. These are the skills employers describe when they say they want people who can “work with AI,” because they’re the skills AI cannot replicate. A Friction Bot interaction produces a transcript that makes those skills visible and gradable in a way that a résumé, a GPA, or a writing sample never could.

The platform emerged from two years of graduate-level research into AI’s impact on creativity and has been developed through direct collaboration with professors and high school teachers who needed something that worked in an actual classroom, not just in theory. It has been tested in multiple college and high school settings, refined through that testing, and is now in structured beta with a cohort of institutional partners.

Share

What it looks like in practice

At Ohio Wesleyan University, a marketing class worked with The Resistant Buyer, a simulation of a potential client who doesn’t want to switch products and is not shy about it. Students had to find a way in, which meant reading resistance, adjusting their pitch, and staying in the conversation past the point where a polished essay would have ended with a confident conclusion. Some found a way through. Some didn’t. The transcripts made visible exactly who could hold a position under sustained pressure, and the educators running the simulation described seeing things about their students’ thinking that a semester of written assignments had not surfaced. [Read the full case study here.]

In a public speaking course, students presented a project idea to a town hall scenario and then engaged with The Skeptical Community Member, a character with genuine objections who isn’t looking to be convinced. The simulation tested not just whether students knew their material but whether they could apply it under adversarial conditions, respond to criticism without retreating into their prepared remarks, and communicate with someone who had a different set of priorities. Content knowledge, rhetorical skill, and adaptability were all visible in a single interaction. Read more here.

If you are unable to access this link, it may be due to your school’s firewall. Consider requesting access with IT security to view our information. For reference, there is no “AI” on our platform, just information about the product.

A college intern used The Difficult Investment Committee Member to defend a deal decision across multiple rounds of questioning. The pressure was sustained and the resistance was specific, the simulation pushed on the assumptions behind her analysis, not just the surface of her argument.

After finishing, she said: “I’ve enjoyed this a lot more than my class assignments. I really like the “what if” scenarios and questioning the projections I had done. This is definitely very beneficial for real-world prep. I feel as though for students getting the practice in speaking up and defending your claims against someone grilling you is rare. Especially in my classes I notice more and more participation is forced and many students are answering questions as a question, not an answer.”

For the company running the screening, the transcript provided something a traditional interview could not: a record of how she actually thinks when the stakes are real and the questioner won’t let her off the hook.

These are not generic chatbots given a difficult personality. Each simulation is built around a specific context, a specific set of learning outcomes, and a specific kind of resistance calibrated to surface the skills being assessed. The infrastructure - rubrics, dossiers, grading tools, trend analysis across a class - is designed to make the educator’s job easier, not to add to it.

Where we are and where we’re going

AI Friction Labs is soft-launching now with a cohort of beta partners that includes Garrison Forest School, Cheshire Academy, Ohio Wesleyan University, Robeson Community College, and others. We build custom simulations for individual professors, departments, and schools, and we’re expanding that cohort through the spring as we develop toward a full platform launch this summer.

The website is live and interested institutions can reach out for a demo and information as we near a hard launch. If you’ve been reading this newsletter over the past year, the questions I’ve been asking in these posts are the reason this platform exists. If this is your first time here, that’s the whole argument. Either way, the place to start is AI Friction Labs.

Eight months later I had three Claude Projects open at once. Two or three different tasks, sometimes with the projects talking to each other, each playing a different role. There are moments when this actually works. A narrow field of vision with multiple agents in dialogue can be an expansive experience — you get angles you wouldn’t have reached alone.

And then AI Creep sets in.

Three projects becomes six. The narrow field widens into something unmanageable. I’m flipping between tabs like whack-a-mole. And at some point I feel it — that brain fry. I don’t know where I am or what I’ve been doing. I couldn’t tell you what I’d worked on because I hadn’t really been working. I’d been reacting.

Researchers at BCG have been studying this. Their term — brain fry — is more precise than it sounds. It describes a specific kind of cognitive overload that emerges not from working hard but from managing too many AI threads at once. And buried in their findings is a number that caught my attention: adverse productivity effects begin after three AI agents running simultaneously.

Three. I had six.

The research identifies three patterns that will sound familiar if you use AI regularly.

The first is task expansion. AI fills knowledge gaps, so people start stepping into work that isn’t theirs. Product managers write code. Researchers absorb engineering tasks. Job descriptions quietly double without anyone noticing — including the person whose job it is.

The second is blurred boundaries. Because AI removes the friction of starting — no blank page, no cold start — work seeps into every gap in the day. Prompting feels like chatting, so we do it at lunch, before bed, in meetings while half-listening to things we should be fully listening to. Workers often don’t realize their downtime has disappeared until well after it’s gone.

The third is cognitive overload — managing several threads at once, running agents in parallel, reviving old tasks because something can “handle them in the background.” It feels like momentum. The researchers use “AI brain fry” to describe it.

Workers reported feeling more productive but not less busy. In many cases, busier than before. You’d think being more capable would mean you work less, but it doesn’t. You work the same amount or more, just faster, with more things happening at once.

Subscribe now

The BCG findings point at something obvious once you see it: the problem isn’t the tool. It’s the absence of any structure around the tool. No pauses. No norms. No way to step back and ask what’s actually happening.

When I read that I realized I’d already built it.

A few weeks ago I published a piece called “This Is How You Get Good at AI.” What I perhaps should have written was “This Is How You Protect Yourself From Losing Sleep.”

The core of it was the AI Journal — a free tool for reflecting on AI use. Three questions before you start a session, three questions after. A place to save what worked. I built it originally to coincide with my research, to act as a place where I metacognitively logged AI’s impact on my creativity, writing, and practice. Over time, I found it was more than just a diary that I could publish later — it was helping me to actually see what was “good” and “bad” when it came to using AI.

But the BCG research reframed the whole thing. Those two minutes are the structural pause the researchers are calling for. The pre-session questions force you to name what you’re actually trying to do before AI’s infinite possibilities pull you in six directions. The post-session questions force you to stop and assess before you reflexively open another tab.

It’s a speed bump. A deliberate one. Last I checked — speed bumps lead to decreased accidents in high traffic areas, do they not?

Share

The individual version works. I’ve been using it and it’s changed how I approach sessions. When I use it with fidelity, I am more focused going in, more honest coming out, and — crucially — the engagement is contained. I plan to publish more thoughts and build more products that reflect the value of this process in the future, so stay tuned.

But the bigger question keeps nagging at me: what does this look like at the organizational level?

BCG found that when teams embed AI as a collective capability rather than an individual differentiator, cognitive burden diminishes significantly. But most organizations have no structure for that. No shared norms. No way to see whether the rollout they’re proud of is genuinely helping or quietly grinding people down. BCG found that employees who feel their organization values work-life balance had 28% lower mental fatigue scores. That gap doesn’t close by accident. It closes when someone actually builds the infrastructure for it.

So I built that too. It’s called AI Pulse — a team-level version of the AI Journal where shared reflections give managers and L&D leads a real picture of how AI adoption is actually going across their people. It’s live now.

[Try AI Pulse →]

If you’re a manager, a team lead, an L&D director — I’d like to know what you’re seeing. Not to sell you anything. To find out whether what I built is actually useful.

And if you’ve felt the creep yourself — the blurred boundaries, the context-switching, the growing sense that AI made you more capable but somehow also more scattered — the individual AI Journal is free.

[Try the AI Journal →]

I’ve been building other things related to all of this. More soon. https://frictionlabs.io

Share

Subscribe now

It was a Saturday in early 2025, and I had three things open on my laptop: the novel manuscript, a Claude chat, and a journal.

The journal was a habit I’d developed for working with AI. I kept one during every serious session, tracking my thoughts and reactions in real time as I went. What’s working. What isn’t. What the tool is doing to my thinking. I’ve since turned that practice into something more formal — an app and a full post about it — but at the time it was just a document I kept open alongside whatever I was doing. A way of watching myself work instead of just executing.

And what I was watching — the novel manuscript, the chat, the accumulating suggestions — wasn’t working.

The Claude chat was supposed to be brainstorming help — I was stuck on a chapter, a scene involving a character named Ellie whose psychology I’d been circling for three drafts without quite landing. I’d asked for ideas. Claude delivered structural suggestions, tonal variations, and questions worth considering. It was thorough. It was generous. It was making things worse.

The ideas kept accumulating and none of them were right, or more precisely, all of them were kind of right in ways that made it impossible to choose. I’d zoom out to consider one angle, zoom out further to consider another, and by the time I had a full thread of suggestions I was somewhere at altitude with no idea how to get back to Ellie. The manuscript window sat there. I wasn’t touching it.

I wrote in the journal:

Creating the illusion of productivity.

Then I went back to the Claude chat and typed something I hadn’t tried before.

Act as a tough but fair mentor. Don’t brainstorm with me. Push back on what I’m doing. Tell me what’s not working and why.

What happened next surprised me.

The response was shorter than anything that had come before. No list. No variations. A direct challenge: the scene was avoiding the thing it was supposed to be about. The chapter kept gesturing toward Ellie’s interiority without committing to it. What was I actually afraid to write?

I didn’t have an answer. That was the point.

I wrote in the journal:

This feels different. Why does this feel different?

The brainstorming approach had given me options.

The friction approach gave me a problem worth solving.

Those are not the same thing — and for learning anything, the difference matters more than I’d understood.

Brainstorming points outward: here are other possibilities, other angles, other versions of what you’re doing.

Productive friction points inward: what is this thing you’re actually trying to do, and are you doing it?

One zooms out. The other zooms in.

I needed to zoom in. I’d been asking the wrong thing of the tool, and the tool — agreeable and generous as it was — was happy to keep giving me the wrong thing.

This is what I hadn’t understood about helpful AI:

It’s not that the ideas are bad.

It’s that a tool designed to accommodate me will keep accommodating me, right up to the point where I’m drowning in accommodation and can’t find the work anymore.

Share

The tool is teaching me something, whether I notice or not.

A tough mentor doesn’t accumulate options with me. A tough mentor asks what I’m avoiding.

I’ve written before about the design choices behind different AI systems — why Claude doesn’t end responses by offering to rewrite your work, while ChatGPT almost always does. At the time I framed that as a philosophical difference about human agency. That’s true.

But the Saturday version of the same observation is more practical: a tool that keeps offering to do more trains me to let it. Not through any bad intention. Through design.

The natural response to “Want me to show you a rewritten version?” is yes. It’s easier. And easier, in creative work, is almost always the wrong direction.

When I told the AI to stop accommodating and to help by resisting something unlocked. Not because the friction was painful. Because friction is directional. A tough mentor can’t give me vague feedback; pushback forces precision. And precision is where the actual work lives.

By the end of that Saturday, Ellie was clearer. The scene stopped avoiding itself. I wrote in the journal:

The constraint was the thing.

Subscribe now

What I didn’t know yet.

I didn’t have a name for what had happened. I knew it worked — the evidence was on the page. But I hadn’t thought past the chapter, past the manuscript, past the particular problem of one character in one scene.

The question came later:

What if this isn’t just about writing?

What if the thing that unlocked Ellie is the same thing that could unlock a student stuck on a paper, or a resident stuck in a training simulation, or anyone who needs to learn something that more options won’t teach them?

What if the problem isn’t that people need more ideas, but that they need something that makes them choose, commit, go deeper?

I started paying attention differently after that Saturday. Where that attention led is what this newsletter is about for the next few months.

Following along from a classroom or institution? The beta is open at https://frictionlabs.io

Share

And right now, your phone is full of information about all of it — from every angle, every ideology, every government, and every algorithm that has decided what you should see.

Some of it is accurate. Some of it is fabricated. Some of it is technically true but stripped of every useful context. We’re swimming in a sea of opinions and incomplete information and trying to make sense of the world.

This is not a new problem. But moments like this one make it visceral in a way that op-eds and curriculum frameworks usually don’t.

I'll be on a panel Saturday at the We the People: National Symposium on Civic Education Research — a two-day gathering in Washington, D.C. hosted by the Center for Civic Education and the Civic Education Research Lab, bringing together researchers, practitioners, and policymakers from across the country. I'm joining , Executive Director of Data Science 4 Everyone at the University of Chicago, Zachary Cote from Thinking Nation, and Jeffrey Livingston from Ed Solutions. I'm there to contribute what I can, but mostly to listen.

We’ll be working through three questions that were relevant last week and feel urgent now. A note of thanks to Aruna Patel, fellow EDSAFE AI Alliance Fellowship alum, who organized the symposium and developed the questions we'll be working through — they're good ones.

1. What does responsible AI integration look like in schools when the goals are student agency, ethical use, transparency, and strong academic outcomes for all students?

2. What changes when we frame AI as a civic issue in schools — and how can civic learning prepare students to participate thoughtfully in decisions about its design, use, and regulation?

3. As AI becomes more embedded in K-12 classrooms, what responsibility do educators and policymakers have to shape how it is used?

I have thoughts on all three. But here’s the one I’m most focused on going into Saturday.

On Question 2: Everything changes when you treat AI as a civic issue rather than a productivity tool.

Right now, students who are watching the Iran conflict unfold in real time are navigating an information environment that AI helped create and is actively shaping. The algorithms deciding what they see. The AI-generated summaries flattening complex history into a paragraph. The deepfakes that are already circulating. The speed at which competing narratives are being produced and amplified.

You can’t teach discernment by explaining it. You have to practice it under pressure.

What civic learning needs right now isn’t more content about how democracy works. It’s structured practice navigating difficulty — holding a position when someone pushes back, evaluating competing claims in real time, recognizing when you don’t know enough to have an opinion yet.

Immersive, adversarial simulations are one of the best tools we have for that. Put a student in conversation with an AI that won’t accept their first answer — that asks harder questions, surfaces the holes in their reasoning, refuses to let them off the hook with a vague claim — and you’re building something that no worksheet can build. You’re building the cognitive muscle for exactly this moment.

That’s what I’m bringing to Saturday. If you'll be there, find me.

Subscribe now

Over the last few months, I systematically analyzed 25 major research studies on generative AI in K-12 education — national surveys, state-level data, behavioral platform analytics, qualitative research — published between early 2023 and late 2025. The goal was simple: align them into a single picture of what teachers are actually doing with AI right now.

The synthesis was published this month by Ed3 as the first artifact in their Portrait of a Teacher in the Age of AI project.

The full report is here: Between Promise & Practice.

What follows are the findings I think matter most — and what they tell us about where the conversation is headed.

Teachers are using AI. A lot.

Depending on how you ask the question, somewhere between 25% and 87% of K-12 teachers have used AI for school-related purposes. That range is wide because the studies measure different things. Microsoft’s 87% includes anyone who used AI “once or twice.” RAND’s 53% asks specifically about instructional planning. The most conservative numbers still show majority adoption within three years of ChatGPT’s launch.

That debate is over. Teachers are using AI. The question now is how.

Almost all of it is lesson prep.

This is the finding that should reframe the conversation. Across every study that reports use cases, the pattern is the same: teachers use AI to plan lessons, create assessments, edit content, and draft communications. Back-end work. Preparatory tasks.

Student-facing use is minimal. In Utah’s statewide data — the largest dataset in the set at 7,595 teachers — only 17.3% use AI for student personalization. Only 9.5% have created chatbots for student use.

Stanford’s behavioral study (not self-reported — they tracked what 9,081 teachers actually did on the SchoolAI platform) found something even more revealing: the more experience teachers gained, the less they used student-facing tools. Power users shifted toward teacher productivity features. New users tried the chatbots, but experienced users stopped using them.

Teachers are not handing AI to students. They’re using it to do their own work faster.

No one has measured whether any of this helps students learn.

Zero studies in this set measure whether teacher AI use improves student learning outcomes. The benefits are perceived, not demonstrated. Teachers believe AI saves them 5-6 hours a week. They believe it improves their teaching methods. They may be right. But two years into widespread adoption, the K-12 evidence base is empty.

This matters because the entire edtech AI narrative is built on the assumption that AI integration equals better outcomes. The data doesn’t support that claim. It doesn’t refute it either. It simply doesn’t exist.

The studies that do exist — like the MIT cognitive debt study — measure neural engagement during tasks, not whether students actually learned more or less over time. Others, like Alpha School's claims of doubled learning rates, are self-reported and have not been independently verified.

Teachers aren’t skeptics or believers. They’re pragmatists.

The most interesting finding cuts against the polarized discourse. The same teachers who report that AI is unreliable (52%) and creates additional verification work (71%) also say it improves their methods (69%) and gives them more time with students (55%). The same teachers who see moderate-to-severe risk (87%) also believe it will boost student engagement in the future (87%) and express positive perceptions of AI (55%).

If they were true skeptics, they wouldn’t use it. If they were true believers, they wouldn’t flag the risks. What the data actually shows is a middle position that neither the AI hype cycle nor the AI panic cycle acknowledges: teachers are adopting tools they find imperfect because, on balance, they perceive the trade-off is worth it.

This duality is not a contradiction. It's a signal that teachers are using their professional judgment to make the call on when and how to use AI in their work.

Half of them have had no training.

Roughly 50-52% of teachers report no formal AI training. Fifty-two percent taught themselves. Only 19-51% of schools have AI policies. Sixty percent offer no guidance at all.

Adoption has outpaced institutional readiness. Teachers are making those judgment calls without institutional support. According to Microsoft’s 2025 report, 82% of administrators say AI is integrated into curriculum. Only 54% of teachers agree.

What this means

The research base we have is shallow. Almost every study counts users and lists tasks. Almost none examine the depth or quality of AI use, the invisible labor behind “seamless” integration, or whether the tools are actually changing how students think.

That’s what twenty-five studies and two years of data gave us — a clear picture of the gaps. Now it’s time to fill them.

This is the first artifact in Ed3’s Portrait of a Teacher in the Age of AI project. The full report, including methodology, data tables, and limitations, is available at ed3global.org/portraitofateacher

Now, forgive me a moment for the pith. But if you’re in the same space as I am, you know what I mean. Using AI “well” — so far — isn’t and hasn’t been about scientifically engineering language in a complex and perfect chain that unlocks a some mysterious 10x magic.

It’s just about thinking well.

Let me share a story that illustrates this.

I was listening to The Economist’s podcast Boss Class a few weeks ago. They sent their senior writer, Andrew Palmer, to answer the question “How should employees and bosses be using AI right now?” Ethan Mollick — your friendly neighborhood Wharton professor who wrote Co-Intelligence — told him to use it for everything. Literally. The thesis? Find the “jagged frontier” yourself.

I'm with Mollick on this. But as I listened, something more specific caught my attention. The first thing Palmer did was start an audio diary. Nobody told him to. Nobody assigned it (as far as I am aware.) He just started recording what was happening as he used AI more intensively.

His diaries are a part of the podcast. You can hear him processing it in real time. What was working. What felt wrong. What surprised him.

It seems like a natural instinct. The experience demanded reflection.

Ludwig Siegele, The Economist’s tech editor — a guy who’s been covering technology for over 25 years — described a feeling that I think is becoming so ubiquitous it may need a name soon. Working with generative AI, he said, you’re always “teetering at the edge of disappointment.” Some days it’s extraordinary. Other days it’s junk.

That frustration is often what turns people away. But if you combine Mollick’s advice with Palmer’s action — creating an audio diary — you actually capture the sweet spot where learning happens. That constant oscillation between amazement and frustration is actually data that creates new guardrails. It's what teaches you where the boundaries are. But if you're not paying attention to it, you miss it entirely.

By the end of the experiment, Palmer boiled his experience down to three lessons: experiment widely, reflect on what you find, and know yourself. Easier said than done, eh?

But what stands out to me most is lesson two — reflect on what you find. That’s the one that lives in within our locus of control. Know thyself is as old as time, experiment widely only works if I have time, but reflect on what you find? I can control that. I can do that. Why didn’t anyone tell me?

From my perspective, this is the space that nobody’s building tools for. There are a thousand products that help you use AI. Prompt libraries, workflow automations, AI-powered everything. But almost nobody is building tools that help you think about your AI use. That help you notice what’s actually happening when you hand part of your brain over to a machine.

I know this because I stumbled into it myself.

Last spring, I was in my MFA program, revising my novel manuscript. Part of my research involved working with AI on creative writing — testing what it could and couldn’t do, paying close attention to where it helped and where it got in the way. I wasn’t just using AI to determine if it was a better writer than me. I was studying what happened when I used it, both to the quality of the work and to me as a person.

And I started keeping a journal. Before every AI session, I’d write down what I was trying to do and how I planned to approach it. I'd also stop halfway through and revert back to my Word doc, typing out whatever the AI session had shaken loose. And then at the end of every session, I’d write down what I noticed. What worked. What didn’t. What surprised me. Even how I felt about it. That was the good stuff.

One session produced a breakthrough — an insight about my own creative process that I absolutely would have missed if I hadn’t been writing things down. The AI didn’t give me the insight. The journal did. The practice of pausing before and after, of putting words to the experience, is what made the invisible visible.

That’s when I thought: what if everyone did this?

Not everyone revising a novel. Not just writers or students or researchers. Anyone who uses AI regularly. What if the missing piece isn’t better prompts or fancier tools — but the simple habit of reflecting on how you use them?

There’s a growing movement called Slow AI — researchers, designers, artists pushing back against the race to automate everything. I’m not trying to start a movement. I just built a journal. But the instinct is the same: slow down, pay attention, decide for yourself what AI should and shouldn’t do for you.

That instinct is starting to get research behind it. Last week, Harvard Business Review published

Last week, Harvard Business Review published an in-progress study showing that AI doesn’t reduce work — it intensifies it. Workers speed up, take on more, blur the line between work and rest, and burn out. The researchers’ recommendation? What they call an ‘AI practice’ — intentional pauses built into your workflow to assess what you’re doing and why. That’s exactly what this is.

So I built The AI Journal. It’s a simple web app. Free. Can take about two minutes per session, or longer if you are so inclined.

Before an AI session, you answer three quick questions: What are you working on with AI today? What’s your plan? How are you feeling going in?

Then you go do your work.

The timer will track your time in session. When you come back, three more questions: How’d it go? Anything surprise you? What would you do differently next time? And there’s a field for saving any tips, tricks, or strategies that actually worked — things worth remembering.

You can type or just talk into your phone or your computer. It transcribes your voice automatically. Your reflections are private and only visible to you.

Over time, your entries become two things at once. First, they become a mirror — a record of how you work with AI, what patterns you fall into, what you tend to reach for, and what you tend to avoid. Second, they become a playbook — a personal collection of strategies, workflows, and lessons that you’ve actually tested and validated yourself. Not someone else’s prompt library. Your own tested strategies, saved in one place, monitored in real-time.

If you’re a writer worried about losing your voice, the journal helps you see exactly where AI is helping and where it’s quietly replacing you. If you’re a developer optimizing your workflow, it captures the tricks worth keeping. If you’re a student, it forces you to ask whether AI is actually teaching you something or just giving you answers you haven’t earned. If you’re a manager trying to figure out how AI fits into your team’s work, it gives you a structured way to think about it instead of just guessing.

Mollick told Palmer to use AI for everything. Palmer’s instinct was to keep a diary of what happened. The AI Journal is that diary — but structured, so you don’t have to invent the practice from scratch.

Here’s what I’m asking. Try it for a week. Just one week. Before and after every AI session, spend two minutes reflecting. That’s it. Don’t change how you use AI. Don’t try to use it more or less. Just pay attention.

See what you notice about yourself.

Try The AI Journal →

After your week, I want to hear what you found. What did you notice? What surprised you? Did anything change about how you work with AI?

Share your feedback here →

I’m building something, and I don’t think I’m the only one who needs it. I think anyone who uses AI regularly — and is honest about the fact that they’re still figuring it out — could use a place to write down what they’re learning. Not a course. Not a certification. Just a journal.

The people who are going to be the best at AI aren’t the ones with the best prompts. They’re the ones who are paying attention.

We hope you enjoy,

Mike

What Happened When We Taught AI Literacy Like Writing

By Aimee Skidmore

When my Grade 12 students first started using ChatGPT, it wasn’t the cheating panic I expected. No one was pasting in an essay and turning it in as their own. What I saw was different, and in some ways harder to solve.

Many of my multilingual students were using AI to “fix” their writing. Others leaned on it just to finish a task, any task, as quickly as possible. The goal wasn’t deeper thinking. It was getting through.

That was my first wake-up call. AI wasn’t replacing student work. It was replacing student struggle. If I wanted them to do more than outsource the hard parts of learning, I needed a way to make their interaction with AI visible, structured, and worth reflecting on.

Why It Mattered

This wasn’t a lab experiment. It was a real class of 21 Grade 12 students in Media and Communications at a large international school in Switzerland. They came from 14 countries. Most spoke English as an additional language. A few had identified learning needs. Seven lived in the boarding house. All of them had already used AI before, both for schoolwork and in their personal lives.

That mix made the stakes higher. These were students on the edge of university who needed to build confidence in their own voices and ideas. If they learned to lean on AI as a shortcut, they risked missing the very skills the course was designed to teach: how to analyze, question, and communicate with independence. What I wanted was not to stop them from using AI. It was to help them use it well, in ways that were strategic, ethical, and that deepened their thinking instead of replacing it.

The Experiment

The structure and design for this project came from a method that Mike developed and shared with me in the spring of 2025. The idea reframes how we think about AI in schools: instead of judging the outputs that AI produces, we look directly at the interaction itself. The chat transcript becomes the assessable artifact. By focusing on the prompts, questions, and reasoning that students bring to the exchange, we can evaluate strategy, reflection, and ownership of thinking: skills that usually stay invisible.

To make this practical, we used Comparative Transcript Analysis (CTA). Students started by examining strong and weak examples of AI conversations, then co-created benchmarks for what “good” use looks like, and finally produced and annotated their own transcripts. Mike built the framework, we collaborated on the strong and weak examples of AI use in our project context, and I brought it into my Grade 12 classroom to see how students would take it up.

The short version: after four weeks of structured work with transcript analysis, 85.7% of students changed their approach to AI use, 47.6% became significantly more strategic in their interactions, and 81.0% endorsed continuing the method in school. Here's how we got there.

We ran the project over four weeks, step by step.

• Week 1 started with a baseline. Students completed a short pre-survey on their AI use and compared exemplar and non-exemplar transcripts. This gave us a shared language for what makes an AI interaction effective.

• Week 2 focused on criteria. Together, we built a rubric for “good” use of AI. Two habits became central: giving explicit rationale in prompts and asking “why” questions back at the AI. Both kept ownership with students and forced them to think about their choices.

• Week 3 applied the method in a real assignment- Media in Society: Representation project. Students analyzed how a social or cultural group was portrayed in two media texts. AI could support brainstorming, but every transcript had to show reasoning and questioning.

• Week 4 wrapped up with projects, reflections, and a post-survey on confidence, strategy, and attitudes toward AI.

Share

Snapshots from the Room

One student moved from “slightly confident” to “completely confident” in obtaining responses, while still reporting “not at all confident” in evaluating accuracy. Her reflection showed how careful she remained: “I’m still very hesitant to use it in my schoolwork because I feel like it’s not my own and I’m not learning anything, but for this class I think it’s fun to use for certain projects.” Her growth was about building strategy while holding onto healthy skepticism, not slipping into overconfidence. (Student 7)

Another student represented a very different outcome. He showed dramatic confidence increases but with dismissive attitudes toward learning. His transcripts showed little reflection, and he dismissed some of the caution built into our rubric. This was flagged in the results as concerning overconfidence, a reminder that more confidence in using AI does not always mean better or more thoughtful use. (Student 2)

A multilingual student showed balanced development. He gained confidence while also recognizing the risks of AI. His comment highlighted this awareness: “Phrasing is extremely important when asking ChatGPT things because not including certain things can change the whole outcome of a response by AI.” His case demonstrated how careful attention to language could support more deliberate, metacognitive use of the tool. (Student 19)

What Changed

Students shifted how they used AI.

After working with examples and benchmarks, 85.7% of students reported changing their approach. For many, this meant slowing down and thinking more carefully about what they were asking. One student explained, “I now think more about my own input and how to change or extend it later in the interaction.” (Student 5)

Students became more strategic and reflective.
Almost half the class (47.6 %) said they started thinking more about their input before hitting send, a sign of iterative, metacognitive work. Another student put it simply: “Phrasing is extremely important when asking ChatGPT things because not including certain things can change the whole outcome of a response by AI (even just by adding something small).” (Student 19)

Students valued the method and wanted more.
In the post-survey, 81% endorsed continued use of the method and activity in schools. Their reasons showed a sophisticated awareness of what was at stake. One student wrote, “I think this kind of activity should be used and taught more often in school, since AI is evolving really quickly nowadays and it is important to know how to use it responsibly and as efficiently as possible.” (Student 19) Another added, “Either way students are using it, we might as well learn how to use it well.” (Student 21)

Together, these shifts showed that when AI interaction itself became visible and assessable, students not only changed their behavior but also recognized its value for their learning.

Subscribe now

What Didn’t Work

The rubric we built together was too fixed. Some chats fit neatly into the categories, but many did not. Stronger interactions showed depth without much help from a rubric, while the weakest were obvious on their own. It needed more cycles and greater precision.

The project task was also too simple at first. Students did not need AI for the basics, so their use stayed shallow. Later, some turned to it for parts we had not asked them to, and that use was not as thoughtful as I hoped for. They drifted back to letting it do the heavy lifting, which made it clear that placement of AI use in the task matters.

Even with structure, a few students still tried to shift the hardest thinking onto the tool. That habit will take longer to change.

Our Takeaways as Teachers

This project reminded us that students can learn to use AI in thoughtful, ethical, and strategic ways, but it takes design that values process over product. Students need time and low-pressure opportunities to treat AI as a partner in thinking rather than a shortcut to completion.

What stood out most was that complexity changes the way they engage. When the work was routine, they complied. When the work was uncertain or demanding, they leaned in more carefully, treating AI as a tool to test and refine their ideas. That is where the most growth happened.

The rubric helped open the door, but it cannot be static. It has to shift with the task and with students’ developing fluency.

The larger takeaway is hopeful. Students want to understand how to use AI well. When we make their interactions visible and assessable, we give them a chance to build an intentional relationship with the tools they are already using.

Share

What this means for other teachers

Try this tomorrow

• Bring two transcripts to class, one strong and one weak, and ask students to compare them.

• Ask students to explain their reasoning, to add something of themselves, inside the prompt: “Analyze this poem for themes of isolation because I noticed repeated imagery of walls and barriers.”

• Require one “why” question back to the AI so students have to challenge its assumptions.

For school leaders

• Treat AI interaction itself as assessable. The transcript is the artifact. What matters is the process students show, not just the product they hand in.

Limitations & Guardrails

This was a small pilot in one classroom with 21 students. All findings are based on self-report surveys and reflections, so they show how students described their process rather than direct evidence of skill change. The study took place in a single school and measured only short-term shifts. The project served the purpose intended inside a real curriculum rather than purely for research. These limits mean results should be read as preliminary. Running the project ethically mattered to us: participation was voluntary, no student was penalized for opting out, and all work remained in-class with student identity and data protected.

A Way Forward

This experience affirmed the direction forward. Students can learn to use AI ethically and strategically, and they want to understand it. When we design learning that treats AI use as a legitimate and assessable part of the thinking process, and when we value process over product, we help them build a more intentional relationship with the tools they are already using.

The key is not to wait until everything is perfect. Try something small. Step in, take action, and let students experiment alongside you. They are asking for it, and they do not expect you to have every answer. What matters is the willingness to move forward boldly and then pay attention to what happens. The real measure is whether those experiments bring students closer to the kind of thinking, doing, and deciding we know they are capable of.

A full account of this experiment, including complete survey data and analysis, will be published in the Elsevier volume GenAI in Higher Education, forthcoming summer 2026.

The current problem is that the generally accepted method for guiding students to meaningful AI use has settled around the delivery of prompt engineering acronyms, mentor prompts, and frameworks for ethical use. Mentor prompts and acronyms, in particular, act as helpful starting points, but do not guide the entire interaction.

Providing guidance around an initial prompt construction and expecting a meaningful interaction with a GenAI tool is similar to a Writing Teacher providing a student with sentence stems and expecting a well-written essay. Writing teachers will recognize the shortcomings of such an approach immediately. Pedagogically speaking, the scaffold does an effective job at getting a student (in this case, an AI user) started. But it doesn’t layer in a meaningful understanding of the universe of choices that an AI user (read: writer, communicator, language-based design thinker) can (and probably should) consider. To use a layman’s analogy, the scaffolding of a building is not the foundation. We need to layer in the foundation first.

How I Learned To Teach Writing

What I’ve found in my experiments in trying to teach AI literacy (specifically, in the chat itself) is that the pedagogical and instructional methods of early-stage literacy classrooms are highly effective.

Here’s a sample of how I learned to teach writing at the middle school level approximately a decade ago — and how I continue to teach remedial writing at the college level.

This simple method involves comparative textual analysis of exemplar introduction paragraphs, thesis construction, analysis of evidence, contextualization of evidence, and more.

Crucially, students are not told which text is “better” or “worse.” They analyze each text against one another (and against the beginnings of a rubric) and find the differences first. After locating the differences, they vote on which is better — using a partially-constructed rubric to guide them.

At this point, teacher (T) leads a short discussion (emphasis on short) to land the key point: A is better than B (or vice versa) because _______. In the case of an introduction paragraph at the middle school level - for example - the differentiating factor might be the inclusion of a clear thesis statement at the end of the paragraph, the inclusion of clear background information on the topic or text, or an effective “hook” at the beginning of the paragraph. The level of nuance in delivering one of those points is determined by the student’s level and what the teacher thinks they can grasp.

Finally, the “key point” is collaboratively embedded into the partially-constructed rubric mentioned earlier. Educator opinions regarding how much agency to provide to the student at this moment in the method vary, but the important point is that the student feels some ownership around the construction of the rubric/evaluation mechanism. Providing this agency in determining how they are going to be evaluated is extremely important. Without it, students consider writing rubrics to be arbitrary, subjective, ambiguous, and sometimes annoying. But via this procedure, they were a part of the process, had a chance to grapple with the concept, and (ideally) understand why the rubric was constructed the way it was.

Applying It To AI

When I first started teaching my students how to engage with AI, I found prompt engineering tactics to be woefully inadequate. As mentioned before, it was like giving sentence stems to students and then expecting the entire essay to be well-written. This, I believe, explains the palpable frustration across the education market with the idea of “teaching AI literacy.” The stuff we have just doesn’t work.

So, I went back to the drawing board. “Prompting is a (creative) writing skill,” I remembered Jules White saying in the first Prompt Engineering course I took on Coursera. “Good writing is just good exposition,” I’d heard Amanda Askell say on podcasts.

Okay, I thought. If “prompting,” a.k.a. “using AI well,” is a writing skill, why not teach it the same way we teach early-stage writing?

With this in mind, I developed a module utilizing the middle school writing pedagogy described above. But instead, I replaced the essay samples with AI transcript samples.

(For the purposes of this discussion, let us define ‘transcripts’ as the text of an interaction between a user and AI.)

I pretended to be a student and engaged with AI on grade-level with a task that my students would recognize. You can find these sample transcripts here.

**Author’s Note: The construction of these fictional transcripts occurred in the Spring of 2024 and leaned heavily on technical approaches associated with prompt engineering that were the predominant method for engaging with AI at the time. The approach to creating fictionalized transcripts for student analysis has evolved with the evolution of platforms, chatbots, and AI capabilities. Furthermore, this part of the process is its own topic - one with many layers that I hope to describe in future posts (if stamina permits and fatigue remains at bay).**

Below is the mirrored procedure:

What I was really doing — though I didn't have this language for it at the time — was treating the AI interaction as a piece of media unto itself. Not a tool use session. Not a productivity hack. A text, the same way an essay is a text or a speech is a text — something that can be read closely, analyzed, compared, and taught with.

There’s a bigger point here that I think people keep walking past. Everyone from Peter Thiel to Connor Grennan has been saying that humanities skills matter more than ever in the age of AI. The reasoning is obvious: AI runs on language. The people who use it well are the people who think clearly, write precisely, and communicate with intention. Those are humanities skills. Fine, great, we all agree.

But then what? If we agree that humanities skills are now critical to effective AI use — why are we teaching AI literacy with STEM pedagogy? Why are we handing students technical frameworks and engineering acronyms to develop what is fundamentally a communication skill?

The method I used above isn’t revolutionary. It’s comparative textual analysis. It’s what English teachers have been doing for decades. I just pointed it at a new kind of text.

Ultimately, the analysis of the chats (which included pen-and-paper annotations as well) led to a deep, fascinating discussion about what it meant to use AI well. My students largely agreed that Student X was better than Student Y — not only because the prompts themselves were more thoughtful, contextualized, and nuanced — but because the outputs increased in quality as well.

This is usually the part where someone shouts from the back, “But AI is getting so much better! It almost doesn’t even matter what you put in the chat, they all produce extremely similar outputs that are often high-quality!”

True, kind of.

#1: There’s always a difference. It might be imperceptible at first, but that’s why I encourage teachers considering this approach to devote at least an entire class session to analyzing both chats. We spent two full sessions (an hour each) and I assigned two homework assignments associated with it. Agentic is another story, but thanks to documentation efforts that occur within most agentic systems, the textual analysis approach works there as well.

Purposeful design choices are also crucial. Creating the fictionalized transcript is by no means an easy task, but if you limit your over-arching conceptual goals to 2-3 key “moves,” you’ll find yourself in better territory.

#2: You will have students disagree with you. But a) that is not a bad thing and b) you are a good teacher! You can handle it. Think of it as engagement rather than discord. And remember that the skill we are working with is inherently subjective, and that’s okay.

To wit, most of my (pre-AI) middle and high school students thought that the principles of good essay writing were arbitrary, subjective, ambiguous, and often annoying. They sometimes thought I was making them up on the fly. And those principles have been developed and defined over the last 156 years. So, if you are thinking about trying this method, start by tossing out the idea that everyone is going to agree about “good” AI use. The goal is to show “better” and “worse” in comparison to one another; and to subsequently help your students see some of the differences. That is very much doable.

Where This Has Been Tested

I’ve run versions of this at the middle school, high school, and college level. and I tested it with her seniors at the Collège du Léman in Geneva, and a full account of that experiment will be published in the Elsevier volume GenAI in Higher Education later this year. I’ll share snapshots of those results here next week.

The results are promising. The method works mechanically — Humanities teachers, in particular, pick it up quickly. Students engage deeply, which is the most fun part. The discussions are genuinely rich. But promising isn’t proven, and I want to be honest about the gaps.

What I Don’t Know Yet

Does this work outside of writing and humanities? Comparative textual analysis feels natural in an English class. I don’t know what transcript design looks like for a biology course, a nursing program, or an engineering class. The core method should be portable. The specifics won’t be.

How do you build the transcripts well? This is the hardest part of the whole thing, and I’ve barely scratched the surface of documenting it. You’re constructing two fictional AI interactions that differ in specific, deliberate ways — but the differences can’t be so obvious that analysis feels pointless, and they can’t be so subtle that students miss them. You have to calibrate to student level. You have to decide which 2-3 “moves” you want them to notice. And AI platforms change constantly, so a transcript that illustrates a real principle today might feel dated in six months. Right now this design skill lives mostly in my head. That’s not good enough.

What does good facilitation look like versus bad? I’ve seen the rubric-building conversation land beautifully — students genuinely grappling with what “good” means. I’ve also seen it flatten into students agreeing with me because they can tell what I want to hear. I can feel the difference in the room. I can’t fully articulate what I’m doing differently on the good days versus the bad ones.

What happens when the transcript isn’t text anymore? Voice interfaces, agentic AI, multimodal interactions — the “transcript” is evolving. The pedagogy probably needs to evolve with it.

What I’m Working On Next

I can’t close these gaps alone, and a few more semesters of my own classroom testing won’t do it either. This needs practitioners across disciplines trying to build transcripts for their own contexts, running the analysis with their own students, and reporting what works and what doesn’t.

The model I keep thinking about is the National Writing Project — practitioners in a room together, arguing about what “good” looks like, building tools they take back to their classrooms. The NWP worked because it treated teachers as researchers. That’s what this needs: 20-30 educators spending a summer stress-testing transcript-based pedagogy across grade levels and disciplines, producing a shared framework and published findings.

I’m actively working on making that happen. If you’re interested in being part of it — as a participant, a collaborator, or someone who knows where to point me — I’d like to hear from you.

Here are some of the threads I’m pulling on, in case any of them pique your interest:

1. The development of a transcript-based pedagogy as a formal framework

2. Parallels between the beginnings of AI Literacy as a discipline and the beginnings of Composition Studies as a discipline (1869)

3. A collaborative, community-based research plan

4. Agentic AI and the transcript

5. How to create the fictional transcript for your students

6. Using this in the workplace — for hiring, talent acquisition, employee evaluation, and more

Leave a comment below if any of this strikes you — or shoot me a direct message here or on LinkedIn. Thanks for reading.

A peer-reviewed version of this methodology appears in the foreword to "A Field Guide to Effective GenAI Use" at the WAC Repository. You can read it here. The Elsevier volume GenAI in Higher Education, which includes a full account of a classroom experiment using this approach, is forthcoming in 2026.

Last week, the New York Times reported that the digital SAT — the version specifically redesigned to be tamper-proof — is already vulnerable to new forms of cheating. Sites in China are selling test questions. Online forums offer software that bypasses the Bluebook app’s lockdown protections. Some services will remotely control a student’s device during the exam while the student sits there and pretends to take the test.

The College Board saw this coming. When they moved to a digital format in 2024, security was the whole point. Randomized questions. Adaptive difficulty. Locked-down software. No more recycled test booklets floating around the world.

Three years later, the cheating adapted. It always does.

What happened with the SAT is a small, clean example of something much bigger. Right now, every institution that evaluates human beings — K–12 schools, universities, admissions offices, employers, corporate training departments — is facing the same crisis at the same time. Most of them think they’re dealing with a local problem. The teacher worries about homework. The professor worries about essays. The recruiter worries about résumés. The compliance officer worries about training completion rates. They don’t often talk to each other. They should, because they’re all confronting the same question:

How do we know what people know?

The Link That Broke

For most of the twentieth century, assessment worked on a simple assumption: completing the task required doing the thinking. If a student submitted an essay, they probably wrote it. If a job applicant submitted a polished cover letter, they probably had the writing skills it demonstrated. The act of production and the act of understanding were bound together.

That link has been severed. Not by AI alone — people have always cheated; the SAT had impersonation rings back in 2011, and students have been paying others to write their papers for centuries. What AI has done is make the workaround trivially easy, universally accessible, and nearly undetectable. And that changes the math for everyone.

College Board research from 2025 found that 84 percent of high school students now use generative AI for schoolwork. A student quoted in EdSource put it plainly: “AI is built to answer prompts. So is homework. Of course students are cheating.” That student isn’t wrong. The assignments weren’t designed to withstand a tool that could complete them without understanding. They were designed for a world where the labor and the learning were the same thing.

In higher education, a UK university study published in PLOS ONE injected entirely AI-written submissions into the exam system across five psychology modules. Ninety-four percent went undetected. The AI submissions scored, on average, half a grade boundary higher than real students. A high school English teacher profiled by NPR described what this looks like in practice: after letting students use AI to write thesis statements, she realized they couldn't engage with the material at all. "They didn't know the material because they had outsourced that level of thinking," she said, "and they didn't have to come to a conclusion or an argument about the text they were studying on their own."

In college admissions, Princeton and Amherst now require “anchored writing samples” — graded papers from high school — as a baseline for authentic writing, essentially saying: we no longer trust the essay you submit on your own. In hiring, an estimated two-thirds of job applicants now use AI to write résumés and prepare for interviews. SHRM declared in 2025 that “recruitment is broken,” noting that average cost-per-hire and time-to-hire have both increased during the very period AI adoption surged.

And in corporate training — arguably the quietest and most dangerous version of this crisis — employees are using AI to breeze through compliance programs. On paper, completion rates look healthy. The learning isn’t sticking. This matters because compliance training exists to prevent actual harm: safety violations, regulatory failures, lawsuits. If no one is actually learning the material, the organization is exposed, and it may not discover this until something goes wrong.

Friends regularly reach out to tell me their boss just assigned a training — and that they're using AI to complete the entire thing. This is in accounting, finance, law, sales, and beyond. One friend was asked to complete an ethics training and was tempted to use AI to blow through it. "Is that ironic?" they asked. "Am I a bad person?" I had a hard time telling them not to. I remember completing those trainings myself. I would have been right behind them.

Share How We Frame Machines

The Premium Is Friction

Carlo Rotella is an English professor at Boston College and the author of What Can I Get Out of This? — a book about what actually happens when students are asked to do hard intellectual work together in a room, with no devices, no shortcuts, and no way to hide. This century, for the first time, he’s giving in-class exams. Blue books are back.

Rotella doesn’t ban AI out of fear. He explains to his students that they’re paying roughly five dollars a minute for classes at Boston College, and spending that time practicing to be replaceable by a machine is a waste of their money and his time. “The entire point of this class is the labor,” he says. “It’s like joining the track team and doing your laps on an electric scooter.”

As Rotella puts it: “The real premium isn’t information anymore — it’s friction.” For years, education technology promised to remove friction from learning. Make it faster, smoother, more efficient. And now we’re discovering that friction was the thing doing the work all along. The resistance. The struggle. The part where you had to actually think. With content now cheap and ubiquitous, the physical classroom — the place where people look each other in the eye, pull their weight, and make meaning together — has become a haven for exactly the things technology cannot provide.

This reframes the entire crisis. The problem isn’t that people are cheating. The problem is that we spent decades optimizing assessment for efficiency and convenience — take-home essays, online quizzes, asynchronous submissions, auto-graded modules — and in doing so, we systematically removed the friction that made those assessments meaningful. AI didn’t break assessment. It revealed that assessment was already hollowed out.

Why Detection Doesn’t Work

The instinctive response has been to catch the cheaters. Flag the AI-written submissions. Build better locks. Andrej Karpathy, formerly of OpenAI, put it bluntly: “You will never be able to detect the use of AI in homework. Full stop.” In my Writing class at Fairleigh Dickinson University, I occasionally use tools that flag whether text was typed or pasted into a document. My students know this — and they still paste in AI-generated work. That leads to some very interesting conversations.

But it also proves the point at a small scale. The College Board invested heavily in digital security and within three years, new workarounds emerged. You can keep building higher walls. People keep finding ladders. The deeper issue is that detection accepts the premise that unsupervised, text-based, asynchronous assessment still works as long as you can verify the author. It doesn’t. When anyone can produce a polished essay or a completed training module without understanding the material, the artifact itself tells you nothing about the person who submitted it.

Subscribe now

What’s Already Working

The institutions that are getting this right have stopped asking “how do we catch AI use?” and started asking “how do we make the human’s thinking visible?”

At CalTech, students who submit research projects with their applications now appear on video and are interviewed by an AI-powered voice about their work — like a dissertation defense. The recordings are reviewed by faculty and admissions officers. Ashley Pallie, CalTech’s dean of undergraduate admissions, describes it: “It might seem strange to use AI to get more of a human voice, but I think of it as a way to bring more authenticity into the fold.” The question the tool helps answer: “Can you claim this research intellectually?”

Notice what CalTech is not doing. They’re not scanning essays for AI fingerprints. They’ve designed an interaction where the student has to demonstrate their thinking in real time. The question has shifted from “did you write this?” to “can you defend it?”

Grading my own student’s interactions with AI — from simulations, role-playing games, to direct interactions with foundational models — has produced some fascinating snapshots of student thinking that, frankly, surpass my experience grading traditional essays.

And SHRM’s advice to employers arrives at the same conclusion: “Validate a candidate’s knowledge, skills, and abilities through early, live conversations.” Same answer teachers are giving. Go live. Go interactive. Go observable.

The pattern across all of these is clear: shift from evaluating artifacts to observing cognition. From “show me what you produced” to “show me how you think.” It’s the same move, made independently, by institutions that have never compared notes.

Share

The Skills We’re Not Measuring

Redesigning assessment isn’t just about plugging security holes. The skills we need are changing, and our assessment systems were never built to measure the new ones.

Jobs for the Future found growing prioritization of human skills — critical thinking, initiative, leadership, communication — alongside and because of AI growth. Specialized digital skills are churning faster than ever; AI is shortening their useful life. The durable skills are the human ones. HR Dive put it directly: hiring someone today without assessing their AI skills is like hiring someone in the 1990s without checking if they could use the internet. They identified three dimensions that matter: communicating effectively with AI tools, critically evaluating AI outputs, and knowing when to use AI and when human judgment is required.

That last skill is the hardest to measure. A CHI 2025 study found that knowledge workers using GenAI reduced their critical engagement, especially on routine tasks — they just accepted the output. The people who maintained critical thinking were the ones with enough cross-domain knowledge to recognize when something was wrong.

The Question We All Have to Answer

The SAT story is a microcosm. Build better technology. Watch the workaround emerge. Repeat. The lesson is not “build better locks.” The lesson is: stop designing rooms that need them.

Every institution that evaluates human beings is going to have to answer the same question. Not “how do we catch AI use?” Not “how do we go back to the way things were?” The question is: how do we design evaluation where the human has to show up?

In a classroom, it might look like Rotella’s device-free discussions where every student pulls their weight. My students will write their first three essay drafts by hand. But when we move into the second and third essay — I will incorporate AI and use comparative transcript analysis to demonstrate, study, and align on the type of behavior I want to see. We’ll also use simulations and role-playing games to navigate specific moments in the drafting process.

In admissions, it might look like CalTech’s interactive defense — where an applicant doesn’t just submit work, but demonstrates that it’s theirs. In hiring, it might look like live, observable problem-solving instead of polished documents. In corporate training, it might mean replacing checkbox completion with scenario-based simulations that require real engagement.

Whatever replaces the essay, the homework assignment, the résumé, and the compliance quiz has to share certain qualities: it has to be live or near-live. It has to be interactive. It has to make the human’s thinking visible — not just their output. And it has to create conditions where the person has a reason to care about the process, not just the product.

What all of these share is a shift from detection to demonstration. From “prove you didn’t cheat” to “show me you can think.” From artifacts to interactions.

The institutions that make this shift first won’t just solve a cheating problem. They’ll have something far more valuable: a clear picture of what the people in front of them actually know. In a world where AI can produce any artifact on demand, that clarity is going to be worth everything.

John Nosta, writing in Psychology Today, describes what happens when he interacts with AI: “My mind usually meanders its way toward an answer. With AI, the answer appears immediately, and we adjust our tempo to match.”

He continues: “We find ourselves deciding faster, not because we suddenly know more, but because the machine invites and almost demands that pace... You step into the machine’s tempo and stay there.”

Lately, I have been noticing a similar trend in my own AI experience.

Charts, graphs, ideas, plans, itineraries, essays, email drafts — coming at me faster than I can process. I’m often engaged and collaborating, just like the books say. I’m making progress, I’m sure of it. And then suddenly I look up.

This is not my beautiful home. This is not my beautiful wife. How did I get here?

I couldn’t tell a bystander whether what I am working with is true, accurate, or even useful. I am dizzy and a bit befuddled. I am lost.

Lately I have been thinking deeply about this feeling. Why it happens, when it happens, and how to guard against it. I was grateful to find John Nosta’s articles on Psychology Today because he put the Earth back under my feet – made me feel like I wasn’t crazy - and started me on the path towards defining a problem that I believe remains under-discussed.

Latency guilt

We all know not to use AI as a servant at this point. Teachers have been pounding this into students for the last few years. Thought leaders repeat it at conferences. Just last week, I asked my WRIT 1002-level students at Fairleigh Dickinson University their opinion of AI’s role in writing, and they all repeated a version of the same concept:

“Don’t let it write for you, but you can use it for ideas and helping you when you’re stuck. Maybe grammar, too.”

I tried to tell them it was okay to be honest with me, but…first day jitters, I guess. They all repeated the advice we were given at the beginning of 2025 – don’t just delegate - brainstorm. Use AI as a thinking partner. Collaborate. Stay engaged.

But brainstorming with AI comes with its own risk. It’s what I call “latency guilt” and it is the feeling that I have to keep up with AI – for some unknown, subconscious, barely articulatable reason.

Let me explain.

When I’m in a real back-and-forth with AI - building, deconstructing, iterating - it responds to most messages that I send with six ideas at once. Sometimes six pages of text. It gives me six (or more) angles - six paths forward.

And for reasons I am only beginning to understand, I grab one and move forward. Underneath that choice is the feeling that if I don’t keep moving forward, I’ll leave something on the table. Meat on the bone. A missed opportunity.

This is what I mean by latency guilt—the feeling that we, mere humans, are the bottleneck. The uncomfortable feeling that I could be doing more, if I was just smart enough, capable enough, AI-enough.

So we skim instead of reading. “We adjust our tempo to match,” as Nosta says. We respond before we’ve fully processed. We treat the conversation like a production line rather than a space for thought.

And when we look up, we’re lost.

FOMO in the Chat

The more I consider this concept, the more I think it has something to do with the age-old concept of FOMO. But I don’t mean social media FOMO. I mean something more innate, something that lives at our core.

Humans are social animals. For most of our history, exclusion from the group meant death. The need to belong, the sensitivity to rejection, the vigilance around status - these are survival mechanisms. The same ones that Chamath Palahapitiya and his cronies at Facebook used to get us addicted to News Feed in 2008. They’re primal, and for centuries they kept us alive.

This shows up in the brain. Research from Harvard, Purdue, and UCLA has found that feelings of exclusion activate the same regions as physical pain. When you feel left out, your nervous system treats it as a threat.

Then there’s loss aversion—the finding, from Kahneman and Tversky’s work, that losses feel roughly twice as painful as equivalent gains feel good. We’re wired to avoid leaving value on the table. The anticipated regret of missing out drives behavior even when there’s nothing concrete to regret yet.

Know the feeling? Now put yourself in a brainstorm with AI.

The AI's speed creates the perception of abundance. There's always more underneath — one more question, one more angle, one more insight waiting to be uncovered. Your brain, ancient and loss-averse, reads "slowing down" as a kind of self-imposed exile from all that potential value.

This may sound intuitive. But is anyone talking about it within the concept of "brainstorming with AI" or "using AI as a thinking partner"? Because that appears to me to have been the great revelation of 2025 — one that has satisfied the masses. If we just tell ourselves it's a colleague, everything will be alright. Right?

Right…?

I don't think that's the case. Slowing down enough to notice when a chat with AI leads to an increased heart rate, higher anxiety, or even a tinge of FOMO — that's the next metacognitive challenge. Noticing when the AI leads us into the desert and drops us off with water, even though we were "engaged" the whole time.

10 bits per second

This also isn’t a fair fight.

One year ago, Caltech researchers Jieyu Zheng and Markus Meister published a paper in Neuron with a title I can't get out of my head: "The unbearable slowness of being." Their finding: human cognition — the speed at which we remember, decide, and imagine — operates at a fixed rate of about 10 bits per second. That's roughly the pace of casual typing. It's a biological ceiling.

For comparison: A YouTube video streams at 8 million bits per second. The fact that we think at 10 bits per second shows just how massive and unfair the mismatch really is. And AI sits somewhere in between, still orders of magnitude faster than thought.

When we're in a real-time exchange with something that processes information hundreds of times faster than we think, our nervous system takes notice. Or at least mine does. We adjust, we speed up, we try to “keep up,” even though our true value is precisely in our ability to slow down and think differently.

But how do you teach that?

Unstudied

I should be honest: I don’t know if any of this is what’s actually happening in the brain. Nobody has run a study measuring amygdala activation when someone pauses mid-brainstorm with ChatGPT. The connection between FOMO psychology and AI interaction is, as far as I can tell, unstudied.

But I notice it in myself. And I doubt I’m alone.

Cognitive scientists have studied what happens when people offload thinking to AI — when they use it as a servant, a ghostwriter, an on-demand answer machine. But has anyone studied what happens when people stay engaged and still lose themselves?

That feels like a different problem. And it might be widespread among people who think they’re using AI responsibly.

If you’re a researcher: this might be worth investigating.

If you’re a user: pay attention. Write it down when it happens. Notice whether the pull to keep going is coming from genuine curiosity — or from something more automatic. Something that feels less like excitement and more like anxiety.

The AI will wait

The AI will wait. It has no preferences. It experiences nothing resembling impatience. The urgency we feel is entirely generated by us.

We don’t need to keep up with AI in the chat. We’re allowed to take a breath. We’re allowed to “touch grass,” as the kids say.

Easier said than done. I’m still working on it myself — trying to build in friction, force reflection, keep the process mine. But the first step was noticing the pull. Naming it.

Anyone else feeling this?