On this page

Synthesis: A replication of Golding et al. (2025) across the same five U.S. universities, surveying 991 undergraduates in October–December 2025 and comparing them with a cohort from two years earlier. Generative AI use roughly doubled on every measure, familiarity rose, and the share of students who call AI use cheating fell by 14 points. The open-ended responses, analyzed with Pathfinder network scaling, show why: students describe AI as an instructional aid, and both users and non-users say whether it counts as cheating depends on how it is used.

Key Findings

  • Use roughly doubled in two years. College use rose from 36% to 81%, high school use from 19% to 60%, and personal use reached 86%. Having a generative-AI app on a device is now the norm at 80%.
  • Cheating became a minority reading. 42% of students said using AI is cheating, down from 56% two years earlier, a 14-point drop. Women were still more likely than men to say so.
  • Students who call it cheating use it anyway. 30.2% of students who consider AI use cheating reported using it. The old correspondence — "it is cheating, so I do not use it" — collapsed from 46% to 12%.
  • Studying replaced writing as the top use. The five most common college tasks are studying, writing, homework, math problems and explaining material. Two years ago, studying was not on the list.
  • Disciplinary differences are large and counter-intuitive. Students used AI in 88% of natural-science courses but 52% of humanities and 36% of social-science courses, with use in natural sciences exceeding humanities after the two-year gap.
  • The word "cheating" is not how users think. The network built from 804 reasons for using AI centers on learning and understanding, with cheating peripheral. The network from 186 reasons for not using it centers on dishonesty, consequences and distrust of the output.

What two years changed

The study repeats the Golding et al. (2025) instrument item for item across the same institutions, which is what makes the comparison more than another cross-sectional snapshot of generative AI use in higher education. Familiarity rose from the low sixes to a mean of 7.25 on a ten-point scale, a medium-to-large difference between cohorts (d = −0.71), and younger students were more familiar than older ones.

Prevalence moved further than familiarity. In 2023, 36% of respondents had used AI in college; in 2025, 81% had. High school use went from 19% to 60%, and 86% reported using a chat AI in their personal life, unrelated to coursework. Having an app installed rose to 80%, with Black students more likely than White students to have one (86% versus 79%) even though White students were more likely to report using AI in college (86% versus 79% and 78%).

The gender gap the first survey found in familiarity did not replicate, which the authors read alongside Pew's February 2026 finding that the gap in use had closed rather than widened.

Where students use it, and what for

Course area is the sharpest split in the data. Students reported using AI in 88% of their natural-science courses, 52% of humanities courses and 36% of social-science courses. The gap survived modeling of repeated measures: in college, use in natural sciences was 1.69 times the rate in humanities and 2.43 times the rate in social sciences, and humanities exceeded social sciences by 1.44 times.

The task list shifted more than the disciplinary pattern. Studying, writing, homework, math problems and explaining material are the top five, and studying's arrival at the top is the change: the earlier survey did not have it in the top five at all. The Pathfinder networks make the same point qualitatively. Among 804 reasons given for using AI, the organizing concepts were study guides, explaining material, checking work and saving time; cheating appeared as a peripheral node rather than a center. Among 186 reasons for not using it, cheating and academic dishonesty were the center, alongside plagiarism, "not own work", getting in trouble and not learning.

What students think cheating is

Perception moved in the same direction as use. On academic integrity, 42% of the sample considered AI use cheating, against 56% two years earlier, and the joint distribution shifted rather than just its margin: "not cheating, and I use it" went from 10% to 51%, while "cheating, and I do not use it" fell from 46% to 12%.

Students who did consider it cheating were not abstainers. 30.2% of them reported using AI, and their open-ended answers carried a conditional construction that appears on both sides of the question — "depends on use" in the cheating network, "depend" in the non-cheating one. The non-cheating network is organized around understanding, step-by-step help and checking answers, with a learn–tool–helpful chain at its center.

What this means for practice

  • Instructors: assume use, not abstention. With 81% of students using AI in college and nearly a third of self-identified cheaters using it anyway, a policy that depends on non-use is a policy that is not being followed.
  • Instructors: target the studying use case deliberately. Students now reach for AI first to study and to have material explained, which is the use most compatible with learning if the task is designed so the explanation has to be applied.
  • Curriculum designers: do not read low social-science use as low interest. The 36% figure tracks task shape — fewer problem sets, more essays and interpretation — so assessment redesign in those disciplines has to start from what the work is, not from the prevalence number.
  • Administrators: replace the cheating binary in policy language. Students already reason conditionally about how AI is used, and a rule written as a prohibition spends credibility on a distinction they do not accept.
  • Faculty developers: treat the 14-point drop in cheating perception as a teaching problem rather than a discipline problem, and build AI literacy around the distinction students already draw. The norm moved while most course designs did not.

Limitations

  • Five universities, one recruitment channel. The sample is 991 undergraduates recruited through psychology participant pools for course credit, 76% women, mean age 19.81, so the prevalence figures describe this population rather than American undergraduates as a whole.
  • Everything is self-report. Social desirability runs in both directions: admitting to cheating and admitting to AI use are both costly, and the two are measured in the same instrument.
  • Repeated cross-section, not a panel. The comparison is between cohorts two years apart, so it cannot show change within individual students.
  • Course-area counts are student estimates. The 88% / 52% / 36% figures come from students' own counts of courses in which they used AI, not from syllabi, logs or observed use.

Citation

Burke, K. C., Lippert, A., Salomon, I., Neuschatz, J. S., Golding, J. M., & Geoghagan, L. A. (2026). The Proliferation of AI among College Students: A Replication and 2-Year Follow-Up. Teaching of Psychology, 1–12.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.