Cheating rates held steady after chatbots arrived, a survey found

The classroom note

A common worry after ChatGPT's release was that students would simply have the chatbot do their homework and cheating rates would spike. Stanford Graduate School of Education reported, in an article dated 31 October 2023, that Challenge Success, the Stanford-affiliated school-reform nonprofit that has surveyed students for more than fifteen years, was not seeing that pattern in its own data. Researchers Denise Pope and Victor Lee told the school's own publication that 60% to 70% of students had consistently reported at least one cheating behaviour in the previous month across the survey's history, and that the share in 2023, after ChatGPT's release, had “stayed about the same or even decreased slightly.”

What the evidence says

The underlying instrument is a repeated, large-scale student self-report survey that Challenge Success says more than 465,000 students have completed since it was developed by Stanford researchers in 2007, and which explicitly asks about AI use alongside academic integrity. Pope's own words in the Stanford piece are careful: “we haven't seen that bear out in our data so far”, a hedge rather than a claim of certainty, and the researchers flagged that self-reported surveys may underreport actual cheating regardless of AI's presence. A follow-up study Challenge Success later described on its own site, covering roughly 4,354 students across six schools about a year and a half after ChatGPT's release, reported the same pattern held: reported prevalence and patterns of academic integrity had not shifted with AI's growing use.

The implementation question

A stable self-reported rate does not mean nothing changed; it may mean the same proportion of students who would have copied a friend's homework or looked up an answer online are now asking a chatbot instead, with the underlying behaviour relabelled rather than eliminated. The instrument was not designed to catch that substitution, because it asks about cheating in general terms, not about which method a student used. A school cannot read this finding as reassurance that its own policies need no update, only as evidence that overall self-reported prevalence has not visibly jumped.

What holds and what fails

The finding holds as a genuine pattern across a large, repeated, methodologically consistent survey run by an organisation with a long track record on the topic, reportedly replicated in the survey's second year of AI-era data. It fails as proof that AI has not changed how students cheat, what tools they use, or how detectable it is; self-report data is vulnerable to underreporting and to shifting norms about what counts as cheating in the first place, both open questions the researchers themselves acknowledge.

• Does “cheating rate unchanged” mean behaviour is unchanged, or that self-report cannot see the change?
• What specific question was asked, and would a student now count AI use as cheating at all?
• Is this survey being used to justify no policy change, when the researchers themselves call for continued monitoring?

A flat line in survey data is a genuinely useful check on early panic, but it is not the same as a clean bill of health.