When AI cheating meets a proctored exam, grades can collapse

A Brown University economics professor suspected widespread AI cheating after a take-home exam average reached 96 percent. When the final moved to an in-person, proctored format, the average dropped to 48.6 percent, matching patterns seen in two large studies on AI use and student performance.

WTF Index IDIOCRACY
◄ Terminator 0 Idiocracy 4 ►

The story shows AI enabling cheating and masking real learning, with supervised exams exposing weaker student skill.

When AI cheating meets a proctored exam, grades can collapse

A striking case at Brown University has become a warning about what can happen when AI tools enter unsupervised coursework. In one economics class, unusually strong take-home exam results were followed by sharply weaker performance once students had to sit for a proctored, in-person final.

The episode is not only about one course. The source also points to two large studies showing the same broad pattern: homework looks better after AI use, while supervised exams tell a much less flattering story.

A Brown exam raised immediate doubts

Roberto Serrano, an economics professor at Brown University, suspected that most of his 86 students had used AI to cheat on a take-home exam. The class average was 96 percent, far above the historical range of 65 and 80 percent.

Serrano tested his concern by running the exam questions through ChatGPT. The answers he received were nearly identical to what many students had submitted. A further clue was the method: many students used a convoluted mathematical proof that ChatGPT also selected, even though a more direct solution was available.

The issue was not simply that students performed well. It was that the performance, the structure of the answers, and the mathematical route taken by many students all pointed in the same direction. For Serrano, the take-home results no longer looked like a reliable measure of what the class had learned.

The final changed the picture

After warning the class, Serrano moved the final exam to an in-person, proctored format. The result was a dramatic reset. Eighteen students dropped the course, and nine did not appear for the test.

The average on the final fell to 48.6 percent, the lowest result the course had ever recorded. Only a small number of students came close to matching their take-home performance, while nineteen students failed outright.

Serrano then voided the midterm and made the final count for 80 percent of the course grade. That decision turned the supervised exam into the central measure of the class outcome, reflecting his view that the earlier take-home result could not be trusted.

The university response frustrated him. According to Serrano, administrators told him to report each suspected cheating case individually, a response he described as "meek" and "Ridiculous." He argued for a stronger institutional stance, saying, "We cannot afford to have a society in which a significant fraction of our best young minds think that cheating is OK," and adding, "That leads to a declining society, to a failed society … We cannot choose to become idiots." Further discussions are ongoing.

Homework gains can hide exam losses

The Brown case fits with a wider concern: AI can make unsupervised work appear stronger without producing the same improvement on supervised tests. Two studies cited in the source show this gap across much larger groups of students.

One study from central China tracked more than 26,000 students in grades 7 through 12 over 30 months. Six months after students began using AI, homework scores increased by 18 percent. At the same time, homework completion time fell from 64 to 45 minutes.

Those gains did not carry over to exams. Exam scores fell by 20 percent. On entrance exams, the long-term loss ranged from 18 to 24 percent, and the full effect took about two years to appear.

The study also found a recurring profile among long-term AI users. About 81 percent showed faster homework completion, high homework grades, and weak exam scores. Top students were affected most, with a 24 percent loss in performance.

Grades rose most where supervision was weakest

A UC Berkeley study covering more than 500,000 grades at a large Texas research university pointed in a similar direction. In courses built around writing and programming assignments, the share of A grades rose 13 percentage points after ChatGPT launched.

The increase was concentrated in unsupervised homework. Courses with a heavy homework component saw an increase of 16 percentage points higher than courses that depended more on proctored exams.

Taken together, the Brown course and the two studies describe a consistent split. AI use appears to make homework faster and more successful on paper, especially when the work is completed outside direct supervision. But when students must demonstrate knowledge under exam conditions, the apparent gains can disappear or reverse.

What the pattern means for education

The facts in the source do not show that every student using AI is cheating, and they do not show that every course faces the same risk. They do show that unsupervised assignments can become a weak signal when AI tools can generate answers, proofs, writing, or code.

For schools, the practical issue is measurement. If homework scores rise while proctored exam scores fall, instructors may be seeing better submissions rather than better understanding. That distinction matters because grades are supposed to reflect learning, not only completed work.

The Brown case also shows how quickly the problem can become institutional. A professor can identify suspicious patterns in one course, but responding student by student is difficult when the concern involves a large share of the class. Serrano's criticism of the university response reflects that tension.

The larger lesson is simple: AI cheating is not only a discipline problem. It can distort the basic feedback loop of education. When take-home work, homework grades, and exam scores no longer point in the same direction, schools have to decide which measures they trust and how they want students to prove what they know.