A new study published in the journal Science warned that the rise of generative AI is pushing universities toward urgent assessment reforms, as traditional coursework becomes increasingly vulnerable to AI-assisted misuse.
The study, called “Generative AI Use and Misuse Call for Assessment Reform in Higher Education,” was conducted by researchers from Cornell University and the University of California, Berkeley, who analyzed survey responses from more than 95,000 undergraduate students across 20 major U.S. public research universities.
It found that 37% of students used generative AI tools such as ChatGPT at least monthly for assignments, while 9% were estimated to have used AI to cheat.
Researchers estimate hidden AI cheating
To estimate AI-assisted cheating, the researchers avoided asking students directly whether they had misused AI. Instead, they split respondents into two groups and asked each student to say only how many statements were true for them, without revealing which ones.
One group saw a standard list of statements about AI use, while the other saw the same list plus one sensitive statement saying they had submitted AI-generated content as their own work while knowing it might not be allowed.
By comparing the two groups’ answers, researchers estimated how common that behavior was without forcing students to admit it directly.
The study found that misuse was more common among frequent AI users. Among students who used GenAI daily, 26% were estimated to have cheated, compared with 7% among monthly users.
AI use spreads unevenly across higher education
Researchers found sharp differences in AI adoption across fields and demographic groups, with regular use highest among computer science students at 62%, compared with 24% among arts students, while 45% of male students reported regular GenAI use versus 33% of female students.
The researchers warned that those gaps could raise equity concerns as AI tools become more advanced, specialized and costly.
Universities face pressure to rethink testing
Rather than calling for blanket bans on AI tools, the researchers argued that universities should redesign assessments around formats that are harder to automate or outsource to AI systems.
The paper pointed to controlled assessments such as oral exams, in-class writing, practical demonstrations and supervised testing environments as stronger ways to verify independent student understanding.
Researchers also highlighted process-based assessment methods that evaluate how students arrive at answers rather than focusing only on final submissions. Those approaches can include requiring students to submit drafts, explain reasoning steps, document AI use, defend conclusions or critique AI-generated outputs.
The study further suggested that assessment reform may need to vary by discipline, since AI affects fields such as computer science, journalism, engineering and the humanities in different ways.
The authors also proposed clearer rules governing acceptable AI use and said professional societies could help define assessment standards for different academic fields.
Rather than treating generative AI solely as a cheating problem, the researchers framed the technology as a broader challenge to how universities measure learning itself.
