A special MIT task force co-chaired by professors Eric Klopfer and Samuel Madden has concluded that AI can now convincingly complete most of the work assigned to undergraduates. In a new report, faculty describe the "deep problems" the technology already creates for education: models produce plausible solutions and reasonable answers to almost any written assignment in the curriculum — essays, math and science problem sets, proofs, programming tasks. Coming from the institution whose degree is a global benchmark for technical rigour, that is an unusually direct statement about what an undergraduate credential currently certifies.
For most instructors, none of this is news. AI-assisted cheating has been visible for years, in some cases escalating to students handing entire remote courses to a model, and universities have been rethinking how they teach and how they test ever since. What MIT has added is an institutional signature on the diagnosis.
The remedies described in the report have a distinctly retro quality. Many professors are moving to oral examinations and handwritten essays. Others want to lean on classroom discussion, require students to keep records of what they have read and studied, or assign more hands-on practical work. This is the pre-digital classroom being reassembled piece by piece, and the choice is revealing: faced with a tool that can imitate the artifact a student produces, the response is to grade things a tool cannot produce yet — presence, speech, handwriting.
The more consequential finding sits elsewhere in the report, and it has nothing to do with cheating. The task force writes that in under three years the technology has measurably changed university life: many students now choose — or feel compelled — to study and solve problems alongside AI. Faculty office hours are less attended. Activity in online course discussions has fallen. Anecdotally, students study together less in dorms, libraries and other shared spaces.
That is the paragraph worth sitting with. Office hours, forums and study groups are not assessment infrastructure, they are the social machinery through which a university actually transmits knowledge — the part that a degree is supposed to be the receipt for. Cheating policy can be rewritten in a semester. A cohort that has stopped asking other humans for help is a harder thing to reverse, and no exam format detects it.
Other institutions are tightening the rules in parallel. The University of Chicago Law School adopted a new "AI strategy" this year, banning phones and laptops in class for first-year courses. Princeton abandoned its Honor Code tradition of more than a century, which had allowed students to sit exams unproctored, after an AI cheating scandal.
Princeton's decision is the one that should make administrators uncomfortable. An honour code is a claim that the institution trusts its students; retiring it is an admission that the claim is no longer underwritten. Every measure on this list — proctors, confiscated laptops, handwriting — buys back the integrity of the grade by spending institutional trust, and there is not an unlimited supply of that.
Notably absent from the public account of the report is any commitment about what MIT will teach, as opposed to how it will examine. The finding is that AI can do nearly any undergraduate assignment; the responses so far are all about making sure the student, rather than the model, produces the assignment. Both can be true at once, and the second does not answer the first. If a proof or a problem set can be generated on demand, the question is not only who typed it but why that task is still the thing being taught.
Universities have spent three years defending the assessment and losing ground every term. The part they have not defended — students walking to office hours, arguing in a library, working a problem badly in front of someone who can see them do it — is the part that was never at risk from a model, only from its convenience.