On 13 August 2026, MIT's Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training published its report. The co-chairs are Eric Klopfer and Sam Madden. In January 2026, Chancellor Melissa Nobles, Provost Anantha Chandrakasan, and Faculty Chair Roger Levy charged the group to assess current AI use at MIT, identify innovations in teaching and student assessment, and propose an AI use policy. The committee writes that what it learned pushed it toward deeper questions about the structure, meaning, and value of an MIT education.
The landscape section says generative AI is already everywhere on campus. Students use it frequently, with mixed feelings. Instructors range from enthusiastic exploration to what the report calls "AI refusal." The authors list effects they find concerning: the technology is upending the p-set, the take-home exam, UROPs, office hours, and the study group, especially for undergraduates; increasing isolation; undermining student mastery and confidence; eroding the "social contract" between instructors and students; making it much harder to assess progress; and challenging MIT norms about rigor, the friction required for learning, collaborative problem solving, and personal integrity. They write that every subject taught at MIT will likely need to be reexamined and potentially revamped so that teaching, content, and assessment are "AI-aware." A drop in in-person study groups is what they say they heard anecdotally.
They start from eight principles. Public engagement with generative AI, they date from ChatGPT in late 2022, so they offer the proposals in humility and expect course corrections. Uncertainty, they write, cannot be an excuse for inaction, and this is not a moment for patches. Humanity comes first: in a listening session, some instructors were considering AI agents as research assistants instead of hiring undergraduates as UROPs, and the committee treats that as a question about what the campus is for. Research at MIT is also an apprenticeship. They want teaching that starts from what students should know, be able to do, and learn to value, using backward design, rather than asking first whether AI is allowed. A poetry seminar, a proof course, and an architecture lab do not need the same rule, and a first-year student is not a doctoral candidate. They borrow "pro-worker AI" from Daron Acemoglu, David Autor, and Simon Johnson and recast it as "pro-learner": AI should expand what students can think about, not replace the hard work of thinking. They name a failure mode they call "cognitive surrender," falling back on a chatbot at the first hint of struggle. Getting the right answer, they write, can create the illusion of learning.
The recommendations open with a claim about the present. Large language models and other generative tools can already produce credible solutions to almost any written assignment in the undergraduate curriculum, including essays, math and science problems, proofs, and coding. Instructors should revisit course goals in every subject before they rebuild assessments. Some older goals may no longer hold; they ask whether most students still need to write complex programs by hand. They urge instructors not to stop at "AI-proofing" the old methods. Many are already weighting in-class exams more heavily or asking students to write or code in the room. The committee says that shift has a cost: it weakens the incentive to do the long p-sets that used to build mastery, and a timed exam is a poor signal if the credential is supposed to mean difficult, independent work. They point instead to oral exams, semester portfolios, and out-of-class assignments paired with in-class conversations.
They want more experiential and project-based work. In the capstone software-engineering class, they write, a semester project often had to be limited and unrealistic; with coding tools, students can now build near production-quality artifacts in a term. Architecture students are using the tools to test ideas faster. The catch, in their wording, is that students still have to stay in charge of the ideas. Because the tools are disrupting social learning, they say every subject should include a regular in-person social component that is not silent note-taking in lecture: group projects with staff check-ins, TA-guided problem solving, rubric discussions, graded in-class talk. UROP, launched in 1969, still reaches 93 percent of undergraduates and 58 percent of faculty. The point of the program, they write, is education, not cheap research labor. If agents become a cheaper stand-in for novice researchers, students lose the relationships through which they learn how a field actually works.
They do not want MIT to ration A's, which they think would intensify the incentive to cut corners. They recommend against relying on AI detectors. The tools may catch purely generated text and miss an outline or an edit; students can answer with "AI humanizers"; false positives can hit non-native English speakers and neurodivergent students; and policing, they write, builds distrust. The Committee on Discipline does not treat detector output alone as enough to bring a case. So-called lockdown browsers should be studied, they say, but the current generation is buggy, error-prone, and feels like surveillance. In-person proctored exams remain the better default for now. They recommend against a single campus AI rule, and they still want every subject to post a clear policy, with a rationale tied to the learning goals. Appendix B offers a four-option menu: unrestricted use, support-tool only, required use, and strictly prohibited. The last, they note, is hard to enforce out of class.
The rest of the report is institutional. They want faster curricular change than year-long multi-committee reviews, an ongoing AI and education committee, school or department AI Leads, instructor support that is not a generic third-party training checkbox, and some accounting of the environmental and financial cost of campus AI use. They also want the residential experience rebuilt around shared in-person work, because a transactional model of assignments as outputs will not stay in the classroom.
The appendices carry the measurements they do have. None of the report text, they say, was generated with AI. Early drafts went to ChatGPT to flag overlapping sections; some graphs and statistical summaries in Appendix C were made with Codex; members also used AI to synthesize other schools' published positions. Their own spring 2026 survey drew 1,632 responses, a 12 percent response rate. ChatGPT was the most used system: 44 percent of respondents said they used it often or very often (46 percent of students, 35 percent of instructors). The spring 2026 Quality of Life Survey, about 8,200 respondents on the main campus, found only 23 percent optimistic about generative AI. Undergraduates reported feeling replaceable at a higher rate than capable (40 percent against 34 percent). A fall 2025 survey by The Tech, MIT's student newspaper, found 90 percent of undergraduates somewhat or very concerned about overreliance on large language models, including 67 percent who were very concerned. Seventy percent agreed that AI proficiency will matter in their careers. Twenty-five percent believed MIT is preparing students to use AI professionally.
The document is a committee report and a menu. It is not a finished Institute policy, and the 12 percent survey is their sample, not a census.