MIT asked a committee for three things: an inventory of AI use, new ideas for grading, and a policy. The committee wrote back that the assignment was too small, because the assignments were already gone. Page nine: “Already these technologies can produce credible solutions and provide reasonable responses to almost any written assignment in our undergraduate curriculum, including essays, math and science problems, proofs, and coding assignments, and their power will only grow.” Lost is my word. The committee’s first line is “This report is a call to action,” and both are true, in that order.
It is the Report of MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, dated August 13, 2026, co-chaired by Eric Klopfer and Sam Madden. Charged in January with those three narrow tasks, the committee says on its first page that what it learned “quickly convinced us” the Institute had to face “deeper questions about the structure, meaning, and value of an MIT education.” In under three years, by its count, attendance at office hours fell, online discussion thinned, and the study group in the dorm started to disappear. “This is not a moment for patches and duct tape.”
Half of what I have been saying for a year is now on MIT letterhead. AI lowered the cost of domain knowledge, and the ivory towers were going to fall with it. The committee just signed the first clause. The next thirty pages are about why the second does not follow, and they convinced me.
The obvious objection is that none of this is new. Solution manuals have circulated for as long as p-sets have, and the point of the p-set was always the work, never the scarcity of the answer. Grant it, because it makes the admission worse. A manual answers the textbook’s question. The model answers yours, at two in the morning, in your own notation. So the crisis starts as an assessment crisis, an instructor unable to tell who did the work, and it does not stay there, because an assignment nobody can verify stops working as practice too. Page 12: AI “saps these tools of their value for both teaching and assessment.” A weekend of reading with a good model at your elbow is earning the domain. A model doing your p-set is renting the answer. The two diverge exactly where the committee says they do: the right answer from a chatbot “can create the illusion of learning” and trigger what it calls “cognitive surrender.” On the Institute’s own survey of about 8,200 people, undergraduates said AI made them feel replaceable more often than capable. The answer got cheap. The knowledge did not, and the report’s stated worst worry, on page five, is the gap: “many uses of AI deprive students of the opportunity to learn.”
Now the honest part, the one a triumphant post about this report always skips. This is no white flag. Past page nine the document is a call to action with a budget attached: oral exams, semester portfolios, take-home work paired with in-class conversation, an in-person social component in every subject, TAs and class time made “more central to evaluation,” a line admitting MIT may have to “find ways to limit class sizes,” and a request to fund an AI Implementation Team, AI Fellows, and a pilot fund. A committee that meant to surrender would not ask for a standing team. Concede all of it, then look at what the plan costs. Every item on that list is a human hour a model cannot supply: a professor across a table, a TA in a smaller room, a graduate student at the next bench. MIT’s answer to cheap answers is expensive humans. That is the admission, and it is stronger for arriving as a budget, because a budget is where an institution says what it cannot do without. One more tell, from the appendix: “None of the text of this report was generated with AI.”
Worth naming where I sit. My company sells agents priced against the salary line of the junior revenue seat. When a professor tells the committee an agent might be cheaper than hiring an undergraduate researcher, that is my customer’s arithmetic one floor up, and a report saying the seat was never about the labor is a convenient report for me. Discount accordingly.
Here is why I read it as true anyway. I did two UROPs at MIT, the program the committee is now worried a research agent will replace, and the report puts its reach at 93 percent of undergraduates. Two summers, two papers: one on combinatorics under Olivier Bernardi, one scoring satellite orbits for Sara Seager’s exoplanet hunt. By the committee’s own sentence, a model could produce a credible version of either today. Which is the committee’s point: the paper was never what the summer was for. “As AI systems become a cheaper or more efficient replacement for UROPs or RAs, students could lose access to the relationships, practices, and shared forms of work through which belonging, confidence, judgment, and professional identity are formed.”
So notice what the plan protects. It keeps the p-set, calling it the ladder to mastery and warning instructors against retreating to timed exams. What it stops trusting is the p-set as proof of work. The money moves toward the struggle the p-set was for, and toward the bench next to the graduate student, the two things a model cannot do on your behalf. MIT’s own committee just drew the line between what can be looked up and what has to be earned, and drew it by admitting the first half is already gone. The answer got cheap. Earning it did not.
Sources:
- Report of MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training (MIT, August 13, 2026)
- Susan Svrluga, “AI can now credibly complete most undergraduate assignments, MIT warns”, The Washington Post, August 25, 2026