A Brown University Professor Made Students Sit a Real Exam. Half the Class Failed.

A professor at Brown University, one of America's eight Ivy League institutions, became so concerned about students submitting AI-generated work that he scrapped the usual assessment format and required a traditional in-person final exam instead. The result was stark: scores fell by roughly 50 percent compared to the coursework grades students had been receiving throughout the term. The story, reported by Ars Technica, has reignited a debate that universities have been having quietly for two years and loudly for the last six months.

The professor's own framing was not subtle. According to the report, he warned that widespread AI cheating leads to what he called 'a failed society', arguing that students are not just gaming a grade but hollowing out the actual skill development that education is supposed to deliver. That is a strong position, but it is not a fringe one. A growing number of academics are reaching the same conclusion: that the gap between AI-assisted coursework performance and genuine understanding has become so large that traditional assessment models are no longer measuring what they claim to measure.

What the 50 Percent Drop Actually Tells Us

A 50 percent decline in scores when external AI tools are removed is a significant data point. It suggests that a meaningful portion of the class was not just using AI to tidy up their writing or check their reasoning, but relying on it to produce work they could not produce themselves. These are students who, on paper, were performing well throughout the semester. In person, without access to a language model, they were not.

It is worth being precise about what this is and is not. It is not evidence that AI cheating makes students stupid. Learning and performance are complicated, and a single high-pressure in-person exam under unfamiliar conditions is not a perfect proxy for genuine understanding either. Some students who use AI tools heavily may still be engaging with the material, using the outputs as a starting point for their own thinking rather than submitting them wholesale. But a 50 percent aggregate drop is too large to attribute to exam nerves or format unfamiliarity alone. Something more systematic is happening.

The incentive structure is the root problem. Universities assess students primarily through submitted work. Submitted work can now be generated in seconds by tools that are free or nearly free to access. The grade for the course reflects the output, not the process. Students who choose not to use AI are, in a meaningful sense, choosing to compete at a disadvantage. That is an unsustainable situation, and it explains why some students who might privately prefer to engage with material genuinely feel structural pressure to reach for the shortcut.

What Universities Are Actually Doing About It

Responses across higher education have been inconsistent and, frankly, often inadequate. Some institutions have leaned into AI detection tools like Turnitin's AI detector, despite those tools having well-documented false positive rates that have incorrectly flagged non-native English speakers at disproportionate rates. Others have issued vague 'AI use policies' that are so hedged as to be nearly meaningless. A smaller number have moved more decisively toward oral examinations, in-class assessments, and project formats that require demonstrated live reasoning.

The Brown case is unusual not because AI cheating is happening there, but because a professor actually tested the hypothesis experimentally and published what he found. Most institutions are operating on suspicion and inference rather than data.

There is also a harder question lurking underneath all of this, one that universities are not yet ready to answer directly. If the skills being assessed are skills that AI can now perform adequately, what exactly is the degree certifying? Some disciplines will have clean answers to that question. A medical student still needs to be able to reason through a diagnosis without a model whispering in their ear. An engineer still needs to understand why a structure fails. But in many humanities and social science courses, the honest answer is more uncomfortable, and the profession is not yet having that conversation at scale.