The policing problem
Faculty say attempts to prevent unauthorized AI use are straining relationships with students. AI-detection tools are unreliable and can wrongly flag work by non-native English speakers and neurodivergent students. The committee advises against using them, warning that they could also fuel a race with “AI humanizers” that disguise machine-written text.
For theses and dissertations, the report proposes disclosure instead: students should explain how they used AI, but should not list it as a co-author.
Students, meanwhile, worry about false accusations and what they see as a double standard. Instructors use AI to prepare slides, write feedback and grade work, while restricting student use. The committee recommends transparency rules for faculty, too.
Exam-proctoring software offers no easy fix. The tools lock down computers and monitor students during tests; the committee says current versions make mistakes and create a sense of surveillance.
What counts as learning
The report’s guiding principle is “augment, don’t automate”: AI should extend human capabilities, not replace them. A correct chatbot answer can feel like proof of understanding, the committee warns, even when the student has not done the work of reaching it. The risk is that students turn to AI at the first sign of difficulty.
That shift could reach MIT’s Undergraduate Research Opportunities Program, or UROP. Faculty are beginning to consider AI agents in place of student research assistants. But UROP is designed to teach students, not provide researchers with cheap labor. At least once, 93% of the class of 2025 took part, and 58% of faculty served as mentors. Replacing students with AI would strip the program of its purpose.
The committee wants instructors to set AI rules course by course, starting with learning goals, then designing assessments, and only then deciding where AI belongs. A poetry seminar and a course in mathematical proofs have different aims; one institute-wide rule could be too permissive for one and too restrictive for the other.
Suggested assessments include oral exams, portfolios built over the semester, in-person discussion and project work. Each course syllabus should state its AI rules and explain the reasoning behind them.
Access, and the evidence behind the warning
Access to paid AI tools could widen differences in student outcomes. MIT provides its community with access to different AI models through Parley, and faculty and graduate students receive $30 a month in free credits. Expanded subscriptions from OpenAI, Google and Anthropic cost several hundred dollars a month, putting them beyond many students’ budgets.
Other research cited in the report points to widespread use and mixed effects:
One finding complicates the report’s case. In a two-year study at Vrije Universiteit Amsterdam, legal scholar Thibault Schrepel randomly assigned students to groups that used no AI, used AI without instruction, or used AI after training. The no-AI group performed best in both years. But the group using AI without instruction also did better, even though students often accepted AI suggestions without checking them. Schrepel had expected the opposite and abandoned his initial assumption that AI helps only when its use is structured.
I think the report’s strongest point is not that AI necessarily lowers learning; the Amsterdam result makes that too simple. It is that universities are changing both instruction and assessment before they can reliably tell what students are learning. The quiet question is how much of a degree can still be measured through work done away from instructors, when AI can supply both the answer and the appearance of competence. Course-level rules may help, but only if the assessments still reveal the difference.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X