A sweeping new study of more than 95,000 undergraduates has found that generative AI use in college classrooms is now widespread enough to threaten the reliability of how universities measure learning—and that a meaningful share of that use crosses the line into outright cheating.
The research, published in Science in May 2026, was led by Igor Chirikov of the University of California, Berkeley, alongside Cornell University’s René Kizilcec and Ivan Smirnov. Drawing on survey data from 20 major public research universities collected during the 2023-24 academic year, the team found that 37% of students used tools like ChatGPT at least monthly to help produce text, code, or video for assignments. Adoption varied sharply by field. Roughly 62% of computer science majors reported regular use, compared with nearly 24% of arts students.
Because students often hesitate to admit rule-breaking even in anonymous surveys, the researchers turned to a statistical workaround called a list experiment, which estimates sensitive behavior without ever asking individuals to confirm it directly. That method put the rate of AI-assisted cheating at roughly 9%—lower than many anecdotal accounts have suggested, though the researchers found cheating climbed to 26% among daily AI users, compared with just 7% among monthly users.
The study also surfaced demographic gaps. Women and students from underrepresented racial minorities reported lower rates of regular AI use than their male, White, and Asian peers, a disparity the authors warn could deepen as AI tools grow more specialized and costly.
Rather than pushing for bans or better detection software, the authors argue for a broader rethinking of assessment. Their recommendations range from a return to closed-book, proctored testing, to clearer classroom rules on acceptable AI use, to redesigning assignments that incorporate AI as a demonstration of professional skill. Because norms differ so much by discipline, the team suggests professional associations—not universities alone—should help set field-specific standards, while remaining alert to unequal access to AI tools.
“Assessment reform is necessary and urgent,” Kizilcec says, warning that unchecked AI misuse undermines “the credibility of university credentials.”
