MIT AI Report Urges Redesign of Grading, Exams and Campus Life

A warning from the university that helped build AI

A committee at the Massachusetts Institute of Technology has concluded that generative AI is changing more than how individual courses work. In its view, the technology is reaching the basic structure of the college experience. MIT's Ad Hoc Committee on AI Use in Teaching, Learning and Research Training was made up of undergraduates, graduate students, faculty and staff, and it was commissioned in January to study how students and instructors use AI12. It found that AI use is everywhere on campus, even though opinions about it differ widely12.

The committee started with a narrow brief: survey current AI use, identify new teaching and assessment practices, and propose a usage policy. Its report says the work soon raised deeper questions about the structure, meaning and value of an MIT education11. MIT President Sally Kornbluth described the moment as a watershed for MIT and for higher education as a whole, the kind of point at which a successful institution has to look again at much of its mission15.

Outlets disagree about when the report came out. One says it was published in June 202613. Public radio coverage describes a faculty-led report released in August15. Wider trade coverage arrived in mid-September12. The differences probably reflect a staged release rather than conflicting facts, but the substance of the report is consistent across all accounts.

What the committee says AI is breaking

The report's main argument is that AI is wearing away the social and teaching structures that MIT's reputation depends on. It lists problem sets, take-home exams, office hours, study groups and the Undergraduate Research Opportunities Program (UROP) among the parts of the experience being disrupted. It also cites growing isolation, weaker student mastery and confidence, and a fraying "social contract" between instructors and students11. The committee says that in under three years, attendance at office hours has fallen, participation in online discussions has dropped, and in-person study groups in dorms and libraries have reportedly thinned out11.

The numbers on student behavior are similar across reports. About 46 percent of surveyed undergraduates said they use large language models every day, while 90 percent said they worry about relying on them too much19. More undergraduates now say AI makes them feel replaceable than say it makes them more capable19. The figures suggest students are uneasy about a habit they have not managed to give up.

The report's term for this is "cognitive surrender," meaning students reach for AI as soon as a problem gets hard instead of working through it. The committee warns that heavy chatbot use can create an "illusion of learning"1216. Chancellor Melissa Nobles connected this to MIT's culture of collaborative problem-solving. She argued that struggle, and even failure, are part of how learning happens, and that AI shortcuts can skip that process15.

UROP is a specific concern. Some faculty are reportedly considering AI agents in place of student research assistants. That matters because 93 percent of the class of 2025 took part in the program at least once, and 58 percent of faculty served as mentors13. The committee recommends that MIT protect and ideally expand UROP and similar mentorship-based experiences11.

The assessment problem and academic integrity

The report's most practical conclusion is that AI has made many familiar assignments unreliable. The committee says AI "saps" problem sets, long written assignments and take-home exams of their value for both teaching and evaluation11. It warns that unrestricted use weakens confidence in grades and credentials11.

The committee does not recommend tighter surveillance. It cautions specifically against AI detectors, and it notes that MIT's Committee on Discipline does not treat a detector's output alone as enough evidence for an integrity case11. The broader trend backs that caution. Yale, Vanderbilt, Johns Hopkins and Indiana have limited faculty reliance on detection software as the sole evidence of cheating, and at least a dozen universities have turned off Turnitin's AI detection feature1. Vanderbilt estimated that even at the vendor's claimed 1 percent false-positive rate, its 75,000 annual submissions could produce roughly 750 wrongly flagged papers3.

The pressure is real. In one national survey, 73 percent of faculty said they had personally handled AI-related integrity cases1. A Harvard Crimson survey found that a third of the graduating class admitted using AI against instructor rules, and 93 percent of those said they were never caught2. Enforcement has costs of its own. Students in California describe proctoring that required mirrors and hands held in view, and a Los Angeles attorney says AI accusations now make up about 35 percent of her firm's education caseload7. Harvard College Dean David Deming has suggested "getting out of the AI-detection business" because it damages trust3.

MIT's alternative is to redesign assessment: oral exams, semester portfolios, in-person discussion and project-based work, with every course stating its AI policy in the syllabus and explaining why13. The committee also warns against swinging too far toward timed, in-class exams. It argues that would narrow what an MIT credential signals and cut back the long projects that build mastery1114. The most notable idea is a thought experiment: if MIT had no grades, much of the incentive to cheat with AI would disappear. The committee explicitly rejects grade rationing but encourages looking at competency-based and mastery-based alternatives11.

One caution about the coverage. A widely shared line saying the "most famous tech school on earth admitted the machines broke its way of teaching" was attributed to the report in at least one outlet19, but it appears to come from social media commentary about the report, not from the document17.

What the tutoring research actually shows

The report does not reject AI. It endorses "augmentation not automation" and notes that AI coding tools now let students in a capstone software course build near-production-quality systems within a semester11. The question is which uses help students learn, and recent research gives conflicting answers.

The optimistic evidence is strong in some places. A Harvard randomized trial in undergraduate physics found that students using a purpose-built AI tutor learned more than twice as much in less time than students in an active-learning class, and reported feeling more engaged28. In K-12 math, randomized trials found that an AI model embedded in live tutoring performed as well as human tutors at fixing errors. Students who used it reached a 66 percent success rate on harder follow-on topics, compared with 61 percent for those tutored by humans alone21.

A newer, larger study points the other way. An October 2026 working paper reports a randomized trial with 2,379 undergraduates and 30 instructors at a large U.S. public university. Students given access to a course-integrated AI tutor had final grades up to 0.37 standard deviations lower and participation in instructor-designed online activities 0.90 standard deviations lower29. The losses for first-generation students were more than twice those of their peers29. The authors also point to earlier findings that unrestricted general-purpose AI can improve performance on assisted homework while hurting independent performance later, a "learning penalty" that tutoring guardrails can reduce29.

Reviews of the field reach similar conclusions. Gains tend to cluster in structured subjects like algebra, and AI tutors often tie or slightly trail well-staffed small-group human tutoring24. Syntheses of generative AI studies report effects that vary so much that their prediction intervals include both benefit and harm28.

The reading: design matters more than the tool

Taken together, the evidence supports MIT's central claim more than either the enthusiasts or the alarmists. AI built to teach, with structure and safeguards, can speed up learning. AI added to a course without changing the course can displace the engagement it was meant to support. The public-university trial's finding that participation collapsed while grades fell29 reads like a measured version of MIT's anecdotes about empty office hours11.

That is why the committee's recommendations focus on institutions more than on tools. They include an ongoing AI and education committee, an AI Pilot Fund for instructors, metrics on AI use and engagement, and norms requiring faculty to disclose their own AI use11. The committee also flags an equity problem: premium AI subscriptions cost hundreds of dollars a month, which could widen performance gaps13. The new tutoring data on first-generation students29 makes that concern harder to dismiss.

The obvious limitation is money. As one commentator notes, the tuition-dependent schools that most need to experiment have the fewest resources to do it20. MIT can pay for extra TAs and redesigned spaces11, and most colleges cannot. Still, the report sets a clear direction for the sector: build assessment around visible thinking and human contact instead of trying to detect AI after the fact. In the committee's words, this "is not an optional exercise"12.