Cornell survey: 95,000 student responses

Cornell and UC Berkeley researchers said on May 21 that generative AI is now embedded deeply enough in undergraduate coursework to raise new questions about how colleges measure learning. Their paper in *Science* drew on 95,513 completed survey responses from students at 20 major public research universities in the United States. The authors said about two-thirds of students used generative AI during the 2023-24 academic year, 37% used it at least monthly, and 9% reported using it to cheat. René Kizilcec of Cornell said the findings point to a problem for “assessment validity” and, by extension, the credibility of university credentials. ### How big was this survey, and what exactly did researchers measure? The study used data from the Student Experience in the Research University, or SERU, Consortium, based at UC Berkeley’s Center for Studies in Higher Education. The supplementary methods say the census-style online survey ran between March and August 2024 across 20 public, research-intensive universities. More than 102,000 students participated overall, and the final analytic sample was narrowed to 95,513 completed undergraduate responses from main campuses. (science.org) Igor Chirikov, a senior researcher at UC Berkeley and a co-author, said the project was meant to replace anecdote with broader evidence about how students use and misuse AI. The paper says earlier studies often relied on smaller samples and weaker ways of measuring sensitive behavior such as cheating, while this one used an indirect method to estimate GenAI-assisted cheating across disciplines. (science.org) ### Where did AI use show up most clearly? The researchers found strong differences by field of study. Cornell said 62% of computer science students reported regular use, compared with 24% of students in the arts. The *Science* paper said GenAI use was concentrated more heavily in disciplines where computational work and data analysis are more central. (science.org) Berkeley’s summary said the same pattern did not carry over neatly to cheating. It said estimated misuse rates were generally higher in non-STEM fields than in STEM fields, with economics and journalism among the higher-rate disciplines and biology among the lower-rate fields. That distinction matters because the paper argues the risks to assessment vary by discipline rather than appearing uniformly across campus. (phys.org) ### What did the study say about cheating and academic integrity? The paper reported that at least 9% of students used AI to cheat. The authors said generative AI can serve as a “cognitive shortcut” that lets students outsource parts of the work that assignments are supposed to evaluate, making some common assessments less reliable as evidence of student capability. (news.berkeley.edu) René Kizilcec, an associate professor of information science at Cornell and director of the Future of Learning Lab, said “assessment reform is necessary and urgent.” He said student misuse of GenAI is “a problem for assessment validity,” and said that creates a problem for the credibility of university credentials. (science.org) ### Did the researchers find gaps in who is using AI? The survey found demographic differences in regular AI use. Cornell said 33% of female students reported regular use, compared with 45% of male students, while students from underrepresented racial minority groups reported lower regular use than white and Asian students. Berkeley’s account also said low-income students used AI less. (news.cornell.edu) The authors said those differences raise concerns about unequal access and unequal practice with AI tools. As GenAI products become more specialized and more expensive, Cornell said the researchers warned that existing gaps could widen. ### What are the authors actually calling universities to do? The *Science* paper says the answer is not a blanket ban or a universal detection regime. (phys.org) Instead, the authors called for discipline-specific assessment reform, arguing that colleges need ways to evaluate knowledge and skills that are harder to fake with AI and better aligned with how students will use AI in professional settings. The paper was published on May 21 in *Science*, and the supplementary materials posted with it set out the survey design, response rates and weighting checks. Kizilcec and Chirikov are listed as corresponding authors, with Ivan Smirnov of the University of Technology Sydney as a co-author. (science.org)