Building the Friction Back In
Relaunching with an essay on what online education can teach the rest of us about AI.
Hello! I’ve been quiet here for a while. Since I last published on this Substack, I’ve been focused on raising a growing family, working at MIT, finishing a graduate degree at Johns Hopkins, writing a novel, and running races like UTMB and the Moab 240.
I’ve had a lot to say about all of this, but I wasn’t ready to write about it yet. I’m back now with fresh ideas for this newsletter. I’ve also merged my separate newsletter, Supply Chain Weekly, here, and there’s a dedicated supply chain tab on the site with selected content that wasn’t in SCW.
The newsletter will remain free, and I’ll be sending three posts a week starting next week. This Wednesday and Sunday are the warm-up. The Friday rhythm will start on June 5th.
Wednesdays will be focused on wearable technologies. (What does your watch know about you, and how secure is that?) Fridays will be focused on robotics security. (Currently, I’m building an AI auditing system.) And Sundays will be for ideas and hard questions on AI ethics and ultra-distance endurance sports.
I wanted to start things off with an analysis of AI in education that bridges some of what I’ve been working on. Let’s dive in!
In an online course I lead on MIT Learn, two groups of students worked on the same assignment in the second week of the course. One group’s average score was 51 percent. The other’s was 68 percent. The students had not been randomly assigned to harder or easier work; they had been randomly assigned to two different versions of an artificial intelligence (AI) tool embedded in the assignment. Same question, different AI partner. Seventeen percentage points.
For most of the past three years, the public conversation about artificial intelligence and higher education has been some version of doom or dismissal. Either ChatGPT will hollow out education credentials and students’ brains, or we will catch the “cheaters” and proceed as usual. That conversation has mostly been about in-person classes, where instructors can take phones away and proctor exams.
Online education does not have the option to pretend AI is not in the room. The substrate is digital, the AI is digital, and the two cannot be compartmentalized. Some might consider this a fatal disadvantage. That small experiment, and those seventeen percentage points, however, suggest that it might actually be an honest opportunity.
In 1962, Douglas Engelbart1 wrote that the purpose of tools like computers was to “augment human intellect… [by] increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems.” He contrasted that vision with the dominant emphasis on automation, and the distinction has been losing the argument ever since. Each generation of digital tools has promised augmentation while steadily shifting human judgment and agency into automated systems. The question now is whether AI is different, or whether we are.
In the course I lead (Supply Chain Technology and Systems), students learn to ask structured questions of a database. (I released a written version of an introductory AI lecture from that course here in the newsletter, with more in the AI series coming.)
In plain English, these questions sound like “what were our total sales last year, broken down by product line, for customers who bought from us at least twice?” Until a few years ago, students wrote out these queries themselves, in a language called SQL. After ChatGPT arrived, I suspect it has been writing SQL for many students.
Within a few months, I saw what I’m sure every instructor in a technical field saw: scores climbing toward the ceiling on questions that hadn’t gotten easier. The grade had stopped measuring what it was meant to measure. (There is a backstory to that pattern — Goodhart’s law — which I wrote about here in the newsletter in the context of Volkswagen’s defeat device.)
The reflex for many in higher education has been to ban the tool. Instead, we built it into the assignment. Two versions, randomly assigned. In one version, the AI was a generator: the student described the query in plain English, and the tool produced the SQL. This interaction is likely similar to how most students use AI tools. In the other version, the AI served as a debugging partner: the student had to write a draft of the query themselves, and the tool would point out errors in their logic or syntax. Both versions sat inside a much harder assignment than we had previously been able to ask — the kind of question we could not assign when students were limited to basic SQL.
The debugger group scored 68 percent. The generator group scored 51 percent. That gap is the gap between augmentation and automation. The generator did most of the cognitive work that the student was meant to learn. The debugger required the student to do the critical work, then helped them improve it the same way a teacher might if they were sitting at the desk with them.
As Dewey put it over a century ago:
Thinking begins in what may fairly enough be called a forked-road situation, a situation which is ambiguous, which presents a dilemma, which proposes alternatives. As long as our activity glides smoothly along from one thing to another, or as long as we permit our imagination to entertain fancies at pleasure, there is no call for reflection. Difficulty or obstruction in the way of reaching a belief brings us, however, to a pause.2
The work of learning lives in friction at the fork. An education that smooths it away has produced something else. The debugger preserved the friction. The generator did not.
Some sources of friction on an assignment are bugs. I run into this constantly when asking students to download and install software for the assignment. But some sources of friction are the lesson. The cognitive scientist Robert Bjork3 called these “desirable difficulties.” Certain kinds of struggle, the ones that feel counterproductive in the moment, are the ones that produce the most durable learning: spacing, retrieval, and the effort of generating an answer before being shown one. An AI tool that smooths those difficulties away has smoothed away the mechanism by which the student was learning.
This is a legitimate reason to remove AI from the classroom, and some research supports the notion. For example, a group from MIT, MassArt, and Wellesley published a study4 last year that used EEG to measure brain activity in students writing essays with ChatGPT, with a search engine, and with no tools at all. The ChatGPT group showed weaker neural engagement, and when the tool was removed, the students could not recall what they had written.
However, there’s also evidence that strategic offloading to an AI partner can enhance rather than erode learning if the freed cognitive capacity gets reinvested in higher-order reasoning. A 2026 study5 of more than 900 students across China, Europe, and the United States found exactly this: when students framed the AI as a partner rather than a shortcut, they activated both critical evaluation of its outputs and strategic delegation of routine tasks. The combination produced deeper learning rather than shallower. A separate report6 from researchers in Australia reached the same conclusion. Offloading is not inherently harmful, but the impact depends entirely on what the learner does with the freed-up cognitive capacity.
This is where online education has an advantage over the in-person classroom. We can measure which redesign preserves the friction. The same digital substrate that makes AI inescapable is the substrate that makes pedagogical science possible at scale. My experiment was run with 526 students.
The principle generalizes outside of technical disciplines. A writing course can put an AI tool inside the assignment, too. One that can offer counterarguments, surface assumptions, and point to sources the student missed, but cannot generate paragraphs on its own. The student submits the essay, the dialogue history with the assistant, and a short note on which of its suggestions they took and which they rejected, and why. The friction has been relocated, not removed. The student is still doing the work of building the argument; the tool is doing the work of pressuring it. I have not run this experiment on a writing assignment yet, so I can only describe the structure, not the result.
The good news is that the work that needs to be done at the instructor level is tractable. Pick one assignment that AI has made trivial (probably most). Allow the student to use AI, but raise the level of difficulty beyond what you could have asked without. And, most importantly, give them instructions on how to use AI. Show them how a debugging partner works and why it will give them better results. If possible, you could also require them to submit evidence of their thinking process, such as the prompt log, a written reasoning trace, or a short oral defense at office hours (for small classes). When you design grading, apply what the writer Carlo Iacono recently called the hard test: if the AI output had been wrong, could the student have caught it and explained why?7
There are important caveats to consider. In a 2026 Brookings Institution report8 on AI, the authors argue that, for young K-12 learners, the risks currently outweigh the benefits. Embedded AI is the wrong move for foundational skills, at any age. A first-year writing seminar is likely not the place to put an AI writing assistant. The introductory tier of a programming language is similarly not the place to embed a code generator.
Additionally, none of this scales on the time of an adjunct teaching five courses for subsistence wages. Embedding AI in the assignment, as our two-version experiment did, requires technical support that most courses may not have access to. This is the current impasse: The instructor cannot wait for the institution, and the institution cannot expect the instructor to do this alone.
Students will arrive next fall with the same tools they had this year. We can turn a blind eye and pretend they are not using it, try to catch them “cheating” when they do, or deliberately build the friction back in. Online education does not get to choose compartmentalization like its in-person cousin, but it can show that the latter option is possible. We can turn the tool into something the student wields rather than something that wields the student.
These are the questions I’ll be writing about here on Sundays. Weekdays will have a technical focus, with behind-the-scenes work on active projects on wearable technologies, robotics, cybersecurity, and AI. Sundays will be for essays like this about AI, ethics, and the demanding environments that test ideas — sometimes the classroom, sometimes the mountains. Next up on Sunday: why I said yes to running two hundred miles in the Italian Alps this fall. If you want to read more before Sunday, two essays are already on the site: the first AI lecture from the course, and a piece on what Volkswagen’s defeat device shows us about deceptive AI systems. If you want the full picture of what this newsletter is working toward, it’s on the About page; otherwise, see you Sunday.
Engelbart, D. C. (1962). Augmenting human intellect: A conceptual framework (SRI Summary Report AFOSR-3223). Stanford Research Institute. https://www.dougengelbart.org/pubs/augment-3906.html
Dewey, J. (1910). How We Think. Boston: D.C. Heath & Co. Available on Project Gutenberg at: https://www.gutenberg.org/files/37423/37423-h/37423-h.htm
Bjork, R.A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe and A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp 185-205). Cambridge, MA: MIT Press. Available at: https://www.researchgate.net/publication/305433736_Memory_and_Meta-memory_Considerations_in_the_Training_of_Human_Beings
Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv. https://doi.org/10.48550/arxiv.2506.08872
Wang, S., & Zhang, H. (2026). Pedagogical partnerships with generative AI in higher education: how dual cognitive pathways paradoxically enable transformative learning. International Journal of Educational Technology in Higher Education, 23(1), 11. https://doi.org/10.1186/s41239-026-00585-x
Lodge J. M. and Loble L (2026). Artificial intelligence, cognitive offloading and implications for education, University of Technology Sydney, doi:10.71741/4pyxmbnjaq.31302475. https://www.uts.edu.au/news/2026/03/experts-warn-unstructured-ai-use-in-schools-risks-cognitive-atrophy/contentassets/ai-cognitive-offloading-and-implications-for-education.pdf
Burns, M., Winthrop, R., Luther, N., Venetis, E., & Karim, R. (2026, January). A new direction for students in an AI world: Prosper, prepare, protect. Center for Universal Education at Brookings. https://www.brookings.edu/articles/a-new-direction-for-students-in-an-ai-world-prosper-prepare-protect/
SUPERB. WOULD YOU CARE TO SHARE WITH OUR CLASS? KUDOS.