On August 25, 2026, MIT released a report from a special committee arguing that the spread of generative AI is changing how students learn and how faculty and students relate to one another. The report calls for a rethinking of courses, assessment methods and research training.
AI can now produce plausible answers to assignments, which makes the traditional approach of judging a student's understanding from the quality of submitted work harder to rely on.
The committee is concerned that AI interactions may start to replace opportunities that students have long relied on: consulting faculty, debating with peers, and learning through repeated failure in the lab.
Rather than simply setting rules for AI use, MIT's recommendations ask the university to first reconsider what it wants students to gain from their education.
How Well Is the Decline in Study Groups Documented?
The report is dated August 13 and was compiled after about five months of discussion and research by undergraduates, graduate students, faculty from each school, and staff from libraries and teaching-support units.
Section 3 states that attendance at office hours and participation in online discussions have fallen, and that in-person study sessions in dormitories and libraries have become less common.
However, the report explicitly says the decline in in-person study groups rests on anecdotal reports gathered through campus interviews.
The survey figures also need to be read with attention to what each one actually measured.
According to Appendix C, a survey on AI use and attitudes conducted by the committee in spring 2026 received responses from 1,632 people, including faculty and students. The response rate, however, was 12%.
This means the possible bias of respondents must be considered; for example, people with a strong interest in AI may have been more likely to respond.
A separate survey of student life, conducted the same spring, drew about 8,200 respondents from the main campus.
Among undergraduates in that survey, 40% said AI had made them feel more replaceable, compared with 34% who said it had improved their own abilities.
These figures reflect how students perceive the situation. They do not measure whether abilities have actually declined or improved. Also, the figure of about 8,200 is the total number of respondents to the survey, not the number of undergraduates alone.
When the report's warnings are sorted by type of evidence, they differ in reliability: the decline in study groups comes from interviews, student anxiety from self-reports, and post-AI academic performance from a separate controlled experiment. They cannot be treated with the same degree of confidence.
| Issue | Type of evidence and subjects | What can be confirmed and its limits |
|---|---|---|
| Decline in in-person study groups | Campus interviews in Section 3 of the MIT report | There are accounts that such gatherings have declined, but the rate of decline and any causal link to AI have not been quantitatively verified |
| Student anxiety | Spring 2026 student life survey in Appendix C.2 | 40% of responding undergraduates said they felt more replaceable. This does not mean their abilities actually declined |
| Performance after AI is removed | Bastani et al.'s high school math experiment | Allows comparison of short-term effects under specific AI and class conditions. It does not prove changes in MIT study groups or university life |
This table organizes the report, its appendices and the experimental paper by their subjects, methods and the extent to which causation can be claimed.
MIT's warning reflects a sense of urgency in the classroom, but the interview findings cannot be recast as scientific proof that AI has isolated students.
The report itself treats loneliness and the decline of in-person interaction as problems that predate AI, and sees AI as adding new pressure to relationships that were already weakening.
AI That Gives Answers vs. AI That Helps Students Solve Problems Themselves
A peer-reviewed paper published in PNAS in 2025 by Hamsa Bastani and colleagues showed experimentally that being able to complete an assignment is not the same as learning.
In fall 2023, researchers ran four math sessions in a Turkish high school with about 1,000 students, assigning classes to a standard conversational AI ("GPT Base"), an AI tutor with teacher-prepared solutions and hints for each problem ("GPT Tutor"), or a control group with no AI.
During practice with AI available, performance was 48% higher with the standard AI and 127% higher with the tutoring AI than in the control group.
But when students were then asked to solve problems without AI, those who had used the standard AI scored 17% lower than the control group.
In Table 1 of the paper, with a perfect score set at 1, the control group averaged 0.321 and the estimated effect of the standard AI was −0.054; dividing this difference by the control mean gives a relative figure of about 17%.
Students who used the tutoring AI, by contrast, showed no statistically significant difference from the control group on the exam taken without AI.
The tutoring AI was designed not to hand over answers but to offer hints and to refer to correct solutions and typical errors that teachers had prepared in advance.
Even with the same GPT-4, the results changed depending on what students were made to do.
However, this study measured short-term effects at a single high school and does not directly show the impact of AI in general today or on long-term learning.
There are also examples where learning improved.
In a Scientific Reports paper by Greg Kestin and colleagues, 194 students in a Harvard physics course experienced, in alternating sessions, a custom AI tutor built on pedagogical principles and in-class active learning.
With the AI tutor, test scores immediately after learning were better than with in-class learning.
This study, too, did not examine long-term retention of knowledge or effects on human relationships. The authors themselves suggest using the AI tutor to prepare for in-person classes.
MIT's principle of "supporting human learning, not doing the learning for students" provides a standard for telling these cases apart.
Even if AI helps students complete an assignment, whether they can explain the content themselves must be checked separately.
And whether an individual's understanding has deepened is a different question from whether students have kept their opportunities to learn with peers and faculty.
Reducing Suspicion Between Faculty and Students
MIT faculty told the committee that the role of policing unauthorized AI use is damaging their relationships with students.
Students, for their part, fear being suspected of using AI when they have not. They are also dissatisfied when faculty use AI for grading and feedback.
If students spend time on an assignment and receive only a mechanical response, the sense that faculty and students are building learning together is likely to weaken.
In Section 3.1.9, the report advises against relying too heavily on AI detection tools.
They have difficulty detecting cases where AI was used only to draft an outline or where parts of the text were edited. They may also wrongly flag writing by students whose first language is not English as AI-generated.
MIT's disciplinary committee likewise does not regard the output of AI detectors alone as sufficient evidence.
As alternatives, the report mentions having students write early drafts in class and requiring assignments to be submitted in several stages.
Instead of examining only the finished product, faculty can check progress along the way, offer advice, and have students explain their thinking.
This allows faculty to evaluate how students thought, and also creates opportunities for students to consult them at the point where they get stuck.
Transparency is also expected of faculty.
The committee recommended that when instructors themselves use AI in course design, materials or feedback, they disclose that use to students.
In Appendix A, the committee itself states that it did not use AI to generate the body of the report, but did use ChatGPT to check an early draft for duplication and Codex to create some charts and statistical summaries in the appendices.
This is one example of clarifying which tasks were handed to AI and who ultimately takes responsibility for the content, rather than asking in blanket terms whether AI was used.
Before AI Agents Replace Research Assistants
The Undergraduate Research Opportunities Program (UROP), which began in 1969, is MIT's research education program for undergraduates.
According to official data for 2024–2025, 93% of the class of 2025 took part in UROP while at MIT, and 58% of faculty served as mentors.
For many students, working in a lab with faculty and graduate students is part of what MIT education means.
However, in the committee's interviews, some faculty were considering using AI agents in place of hiring undergraduates as research assistants.
This does not mean that research assistants are already being replaced on a large scale, or that UROP is being discontinued.
Still, if speed and cost are the only criteria under limited research budgets, opportunities for novices to take part in research could shrink.
Section 2.3 of the report explains that research is central to MIT's mission and is also an apprenticeship-style training ground for the next generation of researchers.
Section 3.1.5 likewise acknowledges the practical value of paid work while emphasizing students' own growth and their relationships with mentors and peers.
Through research work, students learn how to pose questions, how to face failure, and how to make judgments in building knowledge together.
Even if AI finishes the same tasks faster, that alone cannot replace the experience of a student growing as a researcher.
Adopting AI to produce research results quickly and involving students in research to train the next generation of researchers serve different purposes.
The committee urges MIT to maintain UROP and, if possible, expand it.
In deciding whether to replace research assistants with AI, it is necessary to consider not only work time and cost but also who loses the chance to receive mentorship and join a research team.
Design Courses and Assessment Before Setting AI Rules
MIT's recommendations begin by deciding what students should understand and be able to do when they finish a course.
Next comes deciding how to verify that achievement, and only then choosing assignments, class activities and AI-use policies.
A course on building a mathematical proof independently and a course on building large-scale software and verifying its design differ in what can reasonably be delegated to AI.
The usage-policy menu in Appendix B presents the following four types.
| Usage policy | What students are asked to do |
|---|---|
| Free use | AI can be used widely in assignments. Suited to courses centered on assessment without AI, or to assignments where AI use does not undermine the learning goals |
| Limited to support | AI may be used for support such as studying or editing, but not to generate finished answers or most of an assignment |
| Required use as specified | Students use AI with the designated tools and procedures and, where necessary, submit their interactions with the AI |
| Prohibited | Students do not use AI on designated assignments and work on them independently |
This is not a staged scale in which courses move up in order as students gain proficiency.
Different policies can be chosen not only for a whole course but also for individual assignments or parts of them.
Even when supportive use is permitted, unless the policy specifies concretely whether it extends to, for example, drafting outlines or generating figures, faculty and students may understand it differently.
However, these recommendations have not become a uniform rule for all of MIT.
A notice dated August 25 from faculty chair Roger Levy states that for the fall 2026 semester there is no institute-wide requirement for each course to set an AI-use policy.
The menu is presented as a support tool that individual instructors can use as needed.
Nor is the answer on assessment simply to give more weight to in-person exams.
The committee points out that relying heavily on exams could weaken students' motivation to spend time on difficult assignments and projects.
The idea is to combine oral exams, records of work over the semester, out-of-class assignments and in-person checks, so as to assess understanding that is hard to see in short exams alone.
That requires time and support for faculty.
The committee proposes specialized teams and staff to help revise courses, along with a mechanism to fund experimental courses.
The scope includes not only courses that actively use AI but also learning that deliberately avoids it.
It also notes that in courses allowing AI use, a shared environment must be provided so that the AI students can use does not depend on their financial circumstances.
To keep checking whether course changes are actually working, the report calls for tracking AI usage, participation in campus activities, student satisfaction and similar indicators.
Whether the recommendations succeed cannot be judged only by whether the quality of submitted work improves.
Can students explain what they have understood? Can they consult other people? Can they keep opportunities to take part in research settings?
These will serve as the criteria for judging whether the changes align with MIT's educational purpose.