MIT: A Panel Aims to Redefine the Use of AI in Education

A committee of experts from MIT has raised concerns about the erosion of pedagogical interactions and the distrust surrounding the use of AI on campuses. They recommend policies at the course level, more oral and transparent assessment formats, and highlight inequalities in access. Studies conducted at Harvard, Berkeley, Brown, in the UK, and in China report widespread usage and heterogeneous effects on grades.

Studies quantify massive usage and contrasting effects on grades

Several campuses and research efforts describe a broad adoption of AI and divergent impacts on performance. At Harvard in 2024, about 87.5% of respondents reported using these tools, with nearly half using them at least every other day, and around 25% said they were seeking less help during office hours, asking for assistance less often, and skipping readings. In the UK, usage among full-time students reached 95% by the end of 2025, more frequently among affluent households. Anthropic also notes that students delegate analysis and creation tasks to Claude in nearly half of the exchanges studied. Regarding results, a study in China involving over 26,000 students observed an 18% increase in homework grades associated with AI, followed by a 20% drop in exam scores after six months. At UC Berkeley, an analysis of over 500,000 grades found a 13 percentage point increase in the share of A's in writing and coding courses since the launch of ChatGPT, especially where assignments carry significant weight. At Brown, an average of 96% on a take-home exam was followed by a 48.6% on an in-person test. Over two years at Vrije Universiteit Amsterdam, Thibault Schrepel compared three conditions of AI usage: the group without AI finished last over two years, while unstructured usage outperformed the absence of tools, leading him to reject the hypothesis that only structured training makes AI useful.

Detection and lockdown are contested, and transparency is demanded

Teachers report that monitoring unauthorized usage deteriorates relationships with their students. Text generation detectors are deemed unreliable, with false positives targeting non-native or neurodivergent students, and the committee advises against their use, fearing an escalation in response to "humanization" tools. Students fear unfounded accusations and perceive a double standard when teachers use AI for support, feedback, or assessments while limiting its use for learners. Therefore, the panel recommends transparency rules for teachers. Lockdown browsers for exams are described as buggy and akin to surveillance, without addressing the underlying issues. For theses and dissertations, the committee proposes requiring disclosure of any AI usage and banning any co-signature attributed to AI.

AI policies at the course level and rethought assessments

The panel rejects a uniform rule and advocates for policies tailored by course. The definition of these frameworks would start from learning objectives, then assessment design, before specifying what is permitted regarding AI. The committee emphasizes that the needs of a poetry seminar differ from those of a mathematical proofs course. It recommends prioritizing oral exams, semester portfolios, in-person discussions, and projects. Each syllabus should detail the chosen AI policy, with explicit justifications.

Performance gaps related to access to AI tools are reported by the committee

The committee warns that differences in access to AI tools can create significant performance gaps. MIT offers various models accessible via Parley, accompanied by $30 monthly credits for teachers and doctoral students. Meanwhile, the premium plans offered by OpenAI, Google, or Anthropic come with price tags of several hundred dollars per month, an amount considered unaffordable for a large portion of students.

Declining pedagogical interactions and alert on UROP at MIT

The report, published in June 2026, is based on a principle of "augmentation rather than automation" and warns of an illusion of learning related to responses provided by chatbots. At MIT, the committee describes a decline in office hours, study groups, and online interactions, while the assessment of learning outcomes becomes more uncertain for teachers. It warns against the idea of replacing student assistants with AI agents, which would threaten the UROP program. This initiative, presented as a cornerstone of the MIT experience, involved 93% of the Class of 2025 at least once and 58% of faculty as mentors; the panel believes its purpose is education, not substitution by AI. In the background, a gap is noted between the importance students attribute to AI for their careers and their sense of preparedness, and the committee wishes to introduce AI literacy early in introductory courses. It broadly calls on universities to rethink their approach to these tools.