Self-reported cheating among American high school students has stayed about the same since ChatGPT was released, according to a study of the same three schools surveyed before and after November 2022, while the methods have shifted toward AI. A separate blind test at a British university found that 94 percent of exam answers written entirely by AI went undetected by markers. The two findings fit together: cheating is not new, but detecting it by reading the work is hard.
Key takeaways
- Surveys of the same high schools before and after ChatGPT found that overall cheating stayed about level, with AI changing the method rather than the rate.
- In 2024 and 2025, about nine in ten surveyed students did not report having an AI tool write a whole paper, but about half used AI to generate ideas.
- In a blind PLOS ONE experiment, 94 percent of AI-written submissions in five psychology modules were not flagged, and they earned higher average grades than real student work.
- Students and teachers often disagree about which AI uses count as cheating, and some students report being accused when they had not used AI.
Did cheating rise after ChatGPT?
Victor Lee, Denise Pope, Sarah Miles, and Rosalia Zarate analyzed anonymous surveys from three high schools that had collected cheating data both before and after ChatGPT arrived. Their paper in Computers and Education: Artificial Intelligence, published in December 2024, found that cheating behaviors remained relatively stable. Writing in Kappan in December 2025, Lee, Pope, and Miles added data from Challenge Success, a Stanford-affiliated nonprofit whose survey database covers more than 350,000 high school students. In that database, the share of students who said they had paraphrased or copied a few sentences from a source without credit in the past month ranged from 21 percent to 30.1 percent in 2019 and from 24.8 percent to 31.2 percent in 2023. The authors treat that difference as noise.
The authors also cite earlier research in which upwards of 80 percent of high school students reported some cheating behavior. The studies concern U.S. schools, and no comparable before-and-after survey for Canadian schools turned up in this review.
What students say they do with AI
Challenge Success data from the 2024-25 school year show that 49 percent of students used AI to generate ideas for schoolwork and 40 percent used it to revise or improve a paper, project, or assignment. Across 2023-24 and 2024-25, nine in ten students did not report having AI complete a whole paper. The remaining tenth is still large in a class of 30 to 150 students, as the authors point out, and they note that copied papers existed before AI through paper-selling services and classmates.
The harder issue is definition. Spelling and grammar checks are rarely treated as cheating. Asking AI for starter ideas and then developing them may be fine in one classroom and out of bounds in the next, and the authors report that policies differ even between teachers in the same department.
The blind exam test: 94 percent not detected
Peter Scarfe, Kelly Watcham, Alasdair Clarke, and Etienne Roesch submitted work written entirely by AI into five undergraduate psychology modules at a U.K. university, across all years of study, without telling the markers. The PLOS ONE paper, published on June 26, 2024, reports that 94 percent of the AI submissions were undetected. On average they received grades half a grade boundary higher than those of real students, and across modules there was an 83.4 percent chance that an AI submission would beat a random selection of real submissions. The authors frame the finding around the sector's growing reliance on unsupervised take-home exams, where AI use could go unseen.
The result applies to one degree program with a particular assessment style. It does not measure how many students actually submit AI work, only how readily such work passes unnoticed.
Detection errors cut the other way
Students interviewed by the Challenge Success researchers described classmates whose work was flagged as AI-written when they had not used it, including cases where students had simply studied ahead of the course. The authors note that writing by students still developing English proficiency is disproportionately flagged, citing Liang and colleagues (2023), and report hearing of students who ran their own work through AI to avoid being accused. A policy built only on detection therefore risks both missed cases and false accusations.
Three studies at a glance
| Evidence | Setting | What it measured | What it cannot show |
|---|---|---|---|
| Lee and colleagues, 2024 | Three U.S. high schools, before and after ChatGPT | Self-reported cheating stayed relatively stable | Actual rates; three schools only |
| Challenge Success data, reported December 2025 | More than 350,000 U.S. high school students in the database | Nine in ten did not report AI writing a whole paper | Honesty of self-reports |
| Scarfe and colleagues, 2024 | Five U.K. psychology modules, blind submission | 94 percent of AI answers undetected | How many students submit AI work |
Checklist for a course or school AI policy
- State in writing which uses are allowed (brainstorming, grammar checks, feedback) and which are not (submitting generated text as original work).
- Match the rule to the assignment, since the same tool may help one task and defeat the purpose of another.
- Collect process evidence such as drafts, notes, or short oral follow-ups rather than relying on a detector score alone.
- Describe how a student can respond to an accusation before any penalty is applied.
- Review the policy with students, who the Kappan authors say often see AI as a taboo topic with teachers.
FAQ
Is AI making academic dishonesty more common?
The best before-and-after data available found self-reported cheating about level, not higher. Those data come from U.S. high schools, and they rely on students admitting the behavior.
Can teachers tell when work was written by AI?
Not reliably. In the PLOS ONE test, markers did not flag 94 percent of AI-written exam answers, and the answers still scored well.
Is using AI for ideas or editing cheating?
It depends on the instructor's rules, which is why the authors recommend clear, assignment-level policies. Many educators do not count basic spelling and grammar checks as cheating.
What should a student do if wrongly accused?
Keep drafts, notes, and version history, and ask for the school's appeal steps in writing. Institutions differ, so the local academic integrity policy governs the process.
Related reading
- Challenges and limitations of AI in education
- How AI affects students' academic writing
- Challenges and opportunities of AI for education
- Pros and cons of AI in education
More in AI in education.
Last verified: October 2026
Sources
- Cheating in the age of generative AI: A high school survey study of cheating behaviors before and after the release of ChatGPT, Lee V R, Pope D, Miles S, Zarate R C, Computers and Education: Artificial Intelligence 7, December 2024
- Cheating: The AI elephant in the classroom, Lee V R, Pope D, Miles S, Kappan, December 2025
- A real-world test of artificial intelligence infiltration of a university examinations system, Scarfe P, Watcham K, Clarke A, Roesch E, PLOS ONE, June 2024