There is no survey that asks students whether they regret using AI on an assignment, so this side has no measured regret rate. The number used here — 11% — is the cleanest genuine adverse outcome on the action side: a ResumeBuilder survey of 1,000 US job seekers found that 78% got an interview after submitting ChatGPT-written application materials and 59% were hired, but 11% were denied a job once the interviewer realised AI had written the application. The figure most often quoted in this debate — 60% of student AI users (67% of all students) agreeing that AI for schoolwork harms critical thinking, in the RAND American Youth Panel’s December 2025 wave (n=1,214 youth ages 12–29) — is deliberately not used as the rate, because it measures a belief about a general harm, not regret about a specific decision. The tell is that usage kept climbing anyway, from 48% to 62% of students in seven months even as that concern rose from 54% to 67%: students who believe AI is harmful in the abstract overwhelmingly keep using it, which is the opposite of what a regret signal would predict.
The evidence base for actual harm is stronger than the evidence base for actual regret. A field RCT by Hamsa Bastani and colleagues at Wharton, conducted with about 1,000 Turkish high-school math students, found that students given unrestricted ChatGPT-4 access during practice solved 48% more problems correctly during sessions but then scored 17% lower than the control group on a subsequent no-AI exam. The crucial follow-up finding: students did not recognise that they had learned less. If the harm is invisible to the person doing it, self-report regret will systematically understate true regret on a skill-development axis — students who lost skill cannot regret a loss they cannot see. Surveys of college applicants (foundry10, 2024) found about a third had used AI for admissions essays, with 6% having AI write the final draft; 72% of student ChatGPT users in the Intelligent.com survey agreed that using ChatGPT to write college essays counts as cheating, indicating moral discomfort that runs alongside continued use.
The inaction side is genuinely thin, and no published survey gives any percentage for regret among students who wrote assignments themselves. The largest available undergraduate study — the SERU Consortium survey of 95,000+ undergraduates across 20 research universities, published in Science in May 2026 — found that roughly a third of undergraduates do not use AI and that non-use skews lower-income, female, and racially underrepresented, with the researchers warning only, qualitatively and without a number, that these students “may fall behind in college and eventually the workplace.” That framing makes retrospective regret possible among non-users whose abstention is access-constrained rather than principled, but the RAND finding that 78% of non-users believe AI for schoolwork is harmful pulls hard the other way: most students who don’t use AI think they are right not to, which is the opposite of a regret signal. The 10% shown here is therefore a disclosed conservative ceiling for that access-constrained minority, not a measured rate — and because both sides are now anchored on weak, non-equivalent proxies that land close together, the comparison should be read as “no reliable signal that one side dominates,” not as a finding that either choice generates more regret.
The 2023-to-2025 norm shift in this domain has been faster than any other decision in the dataset. ResumeBuilder’s hiring-manager survey in early 2023 found only 18% of hiring managers could identify a ChatGPT cover letter; by 2025, TopResume’s industry data show that roughly a third of hiring managers can spot AI-generated copy in under twenty seconds, 19.6% of recruiters would reject an AI application outright, and 52% accept proofreading or drafting support. Schools have shifted from blanket bans to acceptable-use policies. The figures cited here were taken in one short snapshot of a moving target; the directional finding (action generates more reported concern and concrete adverse outcomes than inaction) is robust, but the magnitudes will move again, plausibly within the next twelve months.