A student submits an elegant essay, a convincing mathematical solution, or a functioning piece of software. The work is better than anything they could have produced a year earlier. Something has improved. What remains uncertain is whether that improvement belongs to the student, the tool, or the partnership between them. Generative AI makes this ambiguity central to education: the quality of a finished product increasingly tells us less about the capabilities of the person submitting it.
An August 2026 MIT committee report confronts this problem with unusual seriousness. Drawing on surveys, listening sessions, and institutional deliberation, it recommends reconsidering assessment, protecting human relationships, and using AI to support intellectual development. Its most consequential argument concerns the distinction between completing educational work and acquiring the abilities that work was designed to develop.
That distinction can disappear in public discussion. A widely circulated X post summarizing the report presented disappearing study groups, declining confidence, and compromised assignments as evidence that machines had broken MIT’s approach to teaching. Many details came from the report, but the narrative was more definitive than its evidence. The document combines observations, concerns, and proposals. It does not establish, through controlled longitudinal research, that AI has caused a general deterioration in MIT students’ abilities.
Its survey findings require similar care. In one undergraduate survey, 46 percent of respondents reported daily use of large language models, while 90 percent expressed some concern about overreliance. These figures describe behaviour and attitudes. Concern about dependency is neither a diagnosis of dependency nor a measurement of lost understanding. The committee’s separate survey had a 12 percent response rate, leaving meaningful uncertainty about whose experiences it captured. These are useful signals for investigation, not a complete account of educational change. MIT’s survey appendix makes their different origins clear.
The underlying concern nevertheless deserves attention. During professional work, delegating a routine task can be sensible. During education, performing that same task may be the mechanism through which a novice develops judgement. An experienced programmer can sometimes delegate implementation because they already understand what to inspect. A beginner may produce working software without developing the knowledge needed to recognise a subtle failure. Identical outputs can therefore conceal very different relationships to the work.
Experimental evidence shows why assistance and learning need separate measures. In a randomised mathematics study, Bastani and colleagues found that access to a general-purpose GPT interface improved students’ performance during practice. When assistance was removed, however, those students performed worse than the control group. A tutor designed with safeguards largely avoided that disadvantage, although it did not produce a corresponding improvement over the control group on the subsequent unaided examination. Successful practice had not reliably translated into independent competence.
Other evidence demonstrates that AI can support learning. In a Harvard physics experiment, Kestin and colleagues found that a carefully structured AI tutor produced higher immediate learning gains than active classroom instruction. The study covered two lessons, so it cannot settle questions about long-term retention or an entire degree. Together, these studies suggest that educational outcomes depend substantially on what the system encourages learners to do: retrieve an answer, attempt an explanation, examine an error, or revise their reasoning.
This also complicates the celebration of difficulty. MIT rightly emphasises the effort involved in developing understanding. But difficulty has different sources. Working through a challenging argument can be productive; struggling with an unclear explanation or waiting days for basic feedback may accomplish little. AI could remove the second kind of friction while preserving, or even strengthening, the first. The educational question is which mental activities must remain with the learner, and which obstacles can usefully be removed.
Assessment must follow that distinction. Oral examinations, portfolios, and discussions of work in progress can reveal things a polished submission conceals. None provides an automatic solution. Portfolios can contain generated material; oral examinations require consistent standards; group projects can hide uneven contributions. A stronger approach combines evidence: what students can explain independently, how they handle unfamiliar problems, and how competently they use and scrutinise AI. Education increasingly needs to assess both independent understanding and effective collaboration with tools.
Time matters as well. Immediate success may reflect temporary support, while genuine learning should remain available later and transfer beyond the original exercise. Institutions should therefore ask what students retain after assistance is withdrawn, what they can apply in a different setting, and whether they can identify plausible errors. Tool-assisted performance remains valuable, but it answers a different question. A student who becomes highly effective with AI may still need opportunities to develop the foundations that make such effectiveness dependable.
MIT’s emphasis on community adds another dimension. Learning includes explaining ideas to peers, receiving criticism, negotiating disagreement, and becoming responsible for shared work. A chatbot may answer a question efficiently while leaving these capacities unpractised. Yet fewer visits to office hours do not, by themselves, prove that learning or belonging has declined. Universities need to examine what replaces those encounters and whether students retain meaningful opportunities for intellectual exchange.
The challenge is to make improved output serve lasting development. That requires deliberate tasks, appropriate assistance, and evidence that students are acquiring capabilities they can understand and exercise. AI may help someone produce an impressive answer before they are ready to defend it. Education succeeds when the support also helps them become ready: able to explain the reasoning, recognise its limits, and carry something learned into the next problem.
No comments yet