AI Cheating in College: Grades Drop From 96% to 48%

What happened

AI dependency in education just produced one of the clearest data points yet: a Brown University class that scored 96 percent on a take-home exam collapsed to an average of 48.6 percent once the same students sat a proctored, in-person final. Economics professor Roberto Serrano grew suspicious when his 86 students posted a class average far above the historical range of 65 to 80 percent. He ran the exam questions through ChatGPT himself and found the chatbot produced nearly identical answers — including an unusually convoluted mathematical proof that many students had also used instead of the more obvious, direct method.

When Serrano switched the final exam to a supervised, in-person format, the results told the real story. Eighteen students dropped the course outright, and nine skipped the exam entirely rather than sit it without AI assistance. Of those who did take it, nineteen failed, and only a handful scored anywhere near their take-home results. Serrano ultimately voided the midterm and reweighted the final to 80 percent of the total grade, calling the university's proposed response — filing individual cheating reports case by case — "meek" and inadequate to the scale of the problem.

Why it matters

Serrano's class isn't an isolated incident. Two much larger studies confirm the same pattern at scale. A 30-month study tracking more than 26,000 students in grades 7 through 12 in central China found that six months after students began using AI tools, homework scores rose by 18 percent while the average time to complete assignments dropped from 64 to 45 minutes. But exam scores fell by 20 percent over the same period, and losses on entrance exams reached 18 to 24 percent, with the full effect taking roughly two years to fully surface. About 81 percent of long-term AI users fit the exact pattern Serrano observed: faster homework, higher homework grades, and weaker exam performance. Top-performing students were hit hardest, losing 24 percent of their exam performance.

A separate UC Berkeley analysis of more than 500,000 grades at a large Texas research university found that the share of A grades in writing- and programming-heavy courses jumped 13 percentage points after ChatGPT's launch — with the effect concentrated almost entirely in unsupervised homework assignments, not supervised coursework. Together, these three data sets point to the same conclusion: AI dependency is quietly decoupling grades from actual learning, and the gap only becomes visible when assessment removes the AI crutch.

How to use it today

None of this means AI tools are the enemy — it means the way they're used determines whether they build skill or mask its absence. Used as a tutor that explains reasoning, checks work, or generates practice problems, AI can accelerate genuine learning. Used as a silent answer-generator for take-home work, it produces exactly the grade inflation Serrano and both studies documented.

The same logic applies well beyond the classroom. Entrepreneurs, marketers, and creators lean on AI daily for drafting, research, and production work, and the healthiest approach is transparency rather than concealment — using AI to speed up execution while still owning the underlying judgment calls. For everyday tasks like transcribing audio, generating images, or converting files, a free toolkit like [MyKreaTool](https://mykreatool.com) can handle the mechanical work quickly, freeing time for the parts of a project that actually require human judgment — the same distinction that separates AI-assisted work from AI-substituted work.

Who benefits

Educators gain the clearest benefit from this data: it's now measurable evidence that take-home, unsupervised assessments are no longer a reliable signal of student ability. Instructors redesigning courses around proctored exams, oral defenses, or in-class writing can point to Serrano's 47.4-point score gap and the Berkeley study's 13-point grade inflation as justification for the change.

Students who are not using AI to cheat also benefit indirectly — inflated grading curves caused by AI-assisted classmates have historically penalized students who did their own work honestly. Restoring proctored assessment resets that curve to reflect actual performance. Employers and university admissions offices benefit too: as grade inflation from unsupervised AI use becomes documented and public, transcripts and homework-based credentials carry less weight, pushing hiring and admissions decisions toward interviews, proctored testing, and portfolio work that can't be quietly outsourced to a chatbot.

Risks

The biggest risk isn't the cheating itself — it's the two-year lag before it becomes visible. The Chinese study found the full academic damage from AI dependency didn't surface for roughly two years after students started relying on it, meaning grade inflation can persist undetected for several semesters before an exam format change or standardized test exposes it. By then, students have often built years of coursework on a foundation of skills they never actually developed.

There's also an institutional risk in how universities respond. Serrano's own administration reportedly wanted individual, case-by-case cheating reports rather than a systemic policy change — a response that treats a structural problem as isolated incidents. That approach risks under-reacting exactly when data from Brown, China, and Texas all point toward the same root cause: unsupervised, AI-assisted assignments no longer measure what they're designed to measure.

Conclusion

The numbers are hard to argue with: a 47.4-point score collapse at Brown, a 20 percent exam-score drop across 26,000 Chinese students, and a 13-point grade inflation jump at a major Texas university all point to the same phenomenon. AI dependency doesn't just risk academic dishonesty — it risks a generation of students, and by extension employees, whose credentials no longer reflect their actual capabilities. The fix isn't banning AI; it's redesigning how skill gets measured, so tools that speed up drafting and research don't end up quietly replacing the underlying learning altogether.