Brown Professor's In-Person Exam Reveals Depth of AI Dependency in Coursework
A Brown University professor, suspecting widespread AI assistance in student assignments, required that the final exam be completed in person without AI access. The result was a roughly 50% drop in scores compared to prior assessments — a data point that has since triggered significant debate across the university and drawn national attention.
The incident is not an isolated complaint about academic integrity. It functions as a controlled test of a question institutions have largely avoided asking directly: how much of current student output is produced by AI, and what remains when that layer is removed?
The answer, at least in this course, was stark.
The core issue is not that students used AI. Many faculty and administrators already assumed that was occurring at some scale. The revelation is the magnitude of the dependency and what it implies about how students are actually engaging with coursework. When the scaffolding was removed, performance did not dip modestly — it collapsed. That suggests AI assistance in this context was not supplementary but load-bearing.
This distinction matters operationally. There is a difference between a student who uses AI to check reasoning, draft a structure, or accelerate research — and one whose submitted work reflects AI output that they cannot reproduce or defend independently. The in-person exam format specifically tested the latter, and a substantial portion of the class failed that test.
For universities, this creates a structural problem with no clean solution. Online assessments, written assignments, and take-home formats are now effectively uncontrolled environments for AI use. Restoring any meaningful baseline requires either moving heavily toward in-person, proctored evaluation — expensive and logistically difficult at scale — or reconceiving what assessment is supposed to measure and how it should be designed for an environment where AI is always available.
The broader implication for employers and professional institutions is arguably more significant. If undergraduate education is producing graduates whose demonstrated competency during training included significant AI assistance they cannot replicate without it, then the credential carries different information than it historically has. Employers hiring for analytical or knowledge-intensive roles may increasingly need their own evaluation protocols to distinguish between candidates who have developed durable skill and those who have learned to produce AI-assisted output that reads as equivalent.
This is not an argument for banning AI from education — that approach is both unenforceable and counterproductive. The more durable question is what baseline competencies need to be verifiable independent of AI, and how institutions structure learning to ensure those competencies are actually developed, not bypassed.
The Brown case is likely to be repeated, with variation, across institutions that begin designing similar tests. What it exposes is that the current academic environment has drifted into an implicit arrangement: AI use is widespread, assessment has not adapted, and the resulting credentials reflect a blend of student capability and AI capability that no one has clearly accounted for. The in-person final did not create the problem — it simply made it visible.
For companies adopting AI in their own operations, the parallel is worth examining. Workflow acceleration through AI is legitimate and often necessary. But organizations that build processes where human judgment is required but humans can no longer perform without AI support — without knowing that is the case — are operating with a hidden dependency that carries its own form of institutional risk.
Sources: — Ars Technica (https://arstechnica.com/ai/2026/07/we-cannot-choose-to-become-idiots-the-ai-cheating-scandal-roiling-brown-university/)