AI Cheating: What the Evidence Says About Students, Detectors, and What Actually Works

Student AI use is now close to universal, detectors are better than they were in 2023, and neither fact settles the question schools are asking. The better question is what happens when a score is wrong, and who pays for it.

Updated October 21, 2026 · 11 min read · Education & AI

Key findings

• Use is near universal. The HEPI/Kortext survey of 1,054 UK full-time undergraduates (fieldwork December 2025, published March 2026) found 95% use AI in some way and 94% use generative AI to help with assessed work. In 2024, the figure for any use was 66%.
• “Using AI” and “cheating” are different measurements. The best pre-2024 behavioral data (Stanford, US high schools) showed overall cheating rates did not rise after ChatGPT launched. That data stops in 2023 and does not cover universities.
• Humans can’t spot it either. In a blind test at the University of Reading, 94% of unedited GPT-4 exam answers went unnoticed by markers. That study tested people, not detectors.
• Detectors have improved, with caveats. An independent University of Chicago Booth benchmark (August 2025) found three leading commercial detectors kept false-positive rates under 1%, with Pangram lowest. Turnitin’s own figures show a gap between document-level and sentence-level error.
• The bias finding is real but dated. The 2023 Stanford study found seven detectors wrongly flagged 61% of non-native English essays. Newer detectors report improvement, but almost all of those figures come from vendors.
• The strongest defense is design, not detection. Process evidence, in-person components, and assessments that assume AI exists.

Table of contents

1. The number everyone skips: behavior vs. tools
2. What changed between 2022 and 2026
3. Why students do it
4. Do AI detectors work in 2026? A source-by-source scorecard
5. The problem that outlasts accuracy gains
6. What institutions are doing instead
7. The cost nobody has measured well: learning
8. Five claims worth retiring
9. What to do: educators
10. What to do: students who get flagged
11. FAQ
12. Sources

The number everyone skips: behavior vs. tools

Most coverage of AI cheating treats late 2022 as year zero: a powerful tool arrived, and integrity collapsed. The behavioral data tells a less dramatic story.

Stanford researchers Victor Lee and Denise Pope have run anonymous surveys at 40 US high schools for years, and added AI questions after ChatGPT launched. About 60% to 70% of students reported at least one cheating behavior in the previous month, and that share was unchanged or slightly lower after ChatGPT arrived. Lee told CNN there was “little evidence that the needle has moved.” Pope’s reading was that most students want to learn and see AI as a way to do it.

The implication is that students who were going to cut corners found a new way to do it, and students who weren’t still aren’t.

Read this finding at its actual size. It is US high school data collected in 2023, based on self-report. It says nothing about 2026, and nothing about universities. Treat it as evidence about the cause of cheating (pressure and incentives, not a sudden collapse in character), not as proof that today’s rates are flat.

What changed between 2022 and 2026

Three things changed, and all three matter more than the headline rate.

Effort collapsed. Buying an essay or leaning on a tutor took money, time, or a favor. A prompt takes seconds.

Normal use became universal. HEPI’s 2026 report describes AI as “near universal” among UK undergraduates. The same survey found institutions lagging: only 36% of students felt encouraged by their institution to use AI, and 38% said their institution provided AI tools.

Detection by humans got hard. In the Reading study, researchers submitted unedited GPT-4 answers under 33 fictitious student identities to a psychology exam system. Markers who didn’t know about the experiment failed to flag 94% of them, and the AI answers averaged about half a grade higher than real students’. The AI did worse in the final-year module. The authors noted real students may also have used AI undetected.

Two details from that study are commonly dropped. First, it tested human markers, not software. Second, the answers were the most detectable kind possible: raw, unedited AI output. Even so, they passed.

Recorded cases rose sharply. A Guardian FOI investigation of 131 UK universities found nearly 7,000 proven AI cheating cases in 2023-24 (5.1 per 1,000 students, up from 1.6), while traditional plagiarism fell from 19 to 15.2 per 1,000. The paper projected about 7.5 per 1,000 for 2024-25.

Be careful with that statistic. These are proven cases, which depend on what institutions can detect and are willing to pursue. Rising recorded cases are consistent with more cheating, with better enforcement, or with a shift in method. The data can’t separate them, and the investigation’s own sources called recorded cases the tip of the iceberg.

Why students do it

Surveys of motive keep returning to the same structural pressures: grades, workload, and assignments that feel like compliance exercises. Stanford’s researchers say as much. That does not excuse cheating, but it changes which fixes will work. If the driver is pressure, a policing response treats the symptom.

HEPI’s 2026 report adds a human detail worth sitting with. Students describe sharply different relationships with the same tools: one says AI helped them focus on deeper analysis, another says they are “not using my brain at all.” A third group reports something quieter: constant worry that honest work will be flagged. One respondent wrote that they worry about detection “even though I have never used AI to write an assignment.”

That last group is the one most detection policies overlook.

Do AI detectors work in 2026? A source-by-source scorecard

“Accuracy” lumps two very different errors together: missing AI text (false negative) and accusing a human (false positive). For a student facing a misconduct hearing, only the second matters. Vendors usually lead with whichever number flatters them, so the table below separates independent from vendor-reported evidence.

Sources: NBER Working Paper 34223 and Chicago Booth Review (Jabarian & Imas, 2025); GPTZero (2026); Turnitin Chief Product Officer update; Higher Ed Dive; Liang et al. (2023). Full links below.

How to read this table. The 2025 Booth benchmark is the strongest independent evidence available, and it is genuinely encouraging: modern commercial detectors, tested on clean samples, make few false accusations. That is a real change from 2023, and an honest article has to say so.

Three limits keep that from being the end of the story:

1. Benchmarks are not your classroom. Lab datasets are not lab reports, discussion posts, or ESL essays written under time pressure. None of the benchmarks above tested the assignment mix of any particular school.
2. Edited text is the hard case. Real cheating is often AI draft plus human editing, or a “humanizer” pass. Detection of mixed or rewritten text is weaker than detection of raw output, and Turnitin added bypasser detection in 2025 precisely because of the arms race.
3. Even 0.5% matters at scale. A 0.5% false-positive rate means one wrongly accused student in every 200 honest submissions. At a university with tens of thousands of submissions a term, that is dozens of people.

The problem that outlasts accuracy gains

The 2023 Stanford study (Liang et al., Patterns) remains the most cited evidence on bias. Seven detectors tested on 91 TOEFL essays written by non-native English speakers misclassified an average of 61.2% as AI-generated. All seven flagged 19.8% unanimously, and at least one detector flagged 97.8%. On essays by US eighth-graders, the same detectors were near-perfect. The authors traced it to perplexity: non-native writers tend to use a narrower range of expressions, which resembles the predictable output of a language model. When the team had ChatGPT enrich the vocabulary of those same essays, false positives dropped, which also shows how easily the tools can be gamed.

What this study is and isn’t:

• It is evidence of a mechanism: detectors that lean on predictability will penalize plain, careful, formulaic prose. That includes many non-native writers and many people trained to write in rigid academic templates.
• It isn’t a measurement of today’s tools. The detectors were 2023 products, the sample was small, and vendors have since retrained. GPTZero reports 1.1% false positives on the same 91-essay set. One third-party summary also notes a separate analysis of roughly 10,000 GRE essays that found no such bias with custom-built detectors. I could not verify that analysis independently.
• It also isn’t evidence of racial bias. The documented effect is tied to language proficiency and writing style.

The fair conclusion in October 2026: the bias problem is probably smaller than in 2023 for the best tools, no independent benchmark has certified it as solved, and almost no vendor publishes a false-positive rate for non-native writers that outsiders can check. Absence of independent verification is itself a reason for caution when the stakes are a failing grade or a misconduct record.

HEPI’s survey points to the human cost. Students told researchers they remain uncertain about what is acceptable, and some fear being falsely accused. A tool can be statistically good and still produce a campus culture of suspicion if it is used as a verdict.

The institutional response has been mixed. Vanderbilt disabled Turnitin’s AI indicator in 2023, citing lack of transparency about how it works. Curtin University announced it would disable the feature from January 1, 2026, keeping text-matching in place. Other universities have kept detection on while restricting how scores can be used. There is no single trend line, so claims like “dozens of universities have banned detectors” deserve a named list before you repeat them.

What institutions are doing instead

The common thread in credible responses is changing what is being assessed, not buying better surveillance.

Secure and open lanes. The University of Sydney’s framework separates “secure” assessments (supervised, in person, AI-free) from “open” ones where AI use is assumed, and the student’s judgment is what’s being evaluated. Each lane has a clear purpose, so students aren’t guessing.

A return to supervised writing. The Wall Street Journal reported in May 2025 that blue-book sales had risen at several large US universities, including UC Berkeley (about 80%), the University of Florida (about 50%), and Texas A&M (about 30%). Those figures come from campus bookstores and are a signal, not a measurement of cheating. In-person writing has real downsides for students who write slowly by hand or who need accommodations. Peter Scarfe, the Reading researcher, has described a wholesale return to sit-down exams as “a step backwards in many ways.”

Process evidence. The HEPI report itself suggests running some AI-free assessments alongside assessments that expect AI use. Drafts, outlines, version history, short oral follow-ups, and annotated reflections on how a tool was used all serve the same purpose: they show thinking, which a score cannot.

Assignments that were already weak. An essay prompt a language model answers well in thirty seconds was often vulnerable to ghostwriters and essay mills before AI. Tying tasks to specific class discussions, local data, or live components makes substitution harder regardless of the tool.

The cost nobody has measured well: learning

There is a risk in the “it’s structural pressure” argument, which is that it can sound like AI use is harmless. The evidence says that depends on how it is used.

HEPI’s 2026 report reflects a divided student body, with some describing deeper learning and others describing disengagement. Coverage of the survey also cites a medical-student case study in which students relying on AI alone, without human feedback, performed worst while feeling the most confident. That pairing, low performance with high confidence, is the more worrying result, and it comes from a small study that deserves replication before anyone builds policy on it.

What the field can’t yet tell you is whether thinking with AI (brainstorming, explanation, critique) and outsourcing to it (drafting the argument) have different long-term effects. Most surveys lump them together. Until that changes, any firm claim about lasting cognitive harm, or about its absence, is ahead of the evidence.

Five claims worth retiring

What to do: educators

Reframe the question. Instead of “how do we catch AI use?”, ask “how would we design this assignment if AI assistance were assumed?” Institutions making progress started there.

This term, do one thing. Redesign one major assignment per course per term. Add a process component (outline, drafts, or a five-minute conversation about the argument) and a short disclosure of any AI use. Treat it as scaffolding, not surveillance.

If you use a detector, set rules before you need them:

• Treat a score as a reason to start a conversation, never as a finding.
• Don’t act on scores under 20% in Turnitin; the vendor says they’re less reliable.
• Require corroboration: draft history, a viva-style follow-up, or marked inconsistencies with the student’s prior work.
• Know your population. If you teach many multilingual writers, a lab benchmark tells you less about your risk.
• Record the tool, the version, and the threshold used, so a challenge can be reviewed.

What to do: students who get flagged

This is general guidance, not legal advice, and processes differ by institution.

1. Ask for specifics. Request the exact score, the tool and version, and which passages were highlighted.
2. Gather process evidence. Google Docs and Word version history, notes, outlines, search history, and earlier drafts are the strongest defense.
3. Point to the vendor’s own limits. Turnitin states that scores under 20% are less reliable and that the indicator shouldn’t be the only evidence in a case. Cite Liang et al. (2023) if language background is relevant.
4. Ask for human review of the writing itself, ideally including a conversation about the content.
5. Keep records going forward. Save drafts as you work. It costs nothing and is the most effective protection.

Writing patterns that can raise false-flag risk include formulaic structure, short passages, heavy automated grammar smoothing, and low-variation technical prose. None of these is evidence of anything, and no one should change their natural writing out of fear.

FAQ

Is AI cheating increasing?

Recorded AI cheating cases rose sharply in the UK between 2022-23 and 2023-24, according to a Guardian FOI analysis. Overall self-reported cheating among US high schoolers did not rise through 2023, according to Stanford’s surveys. Both can be true, because the first measures detected cases and the second measures reported behavior. No current dataset cleanly measures overall rates in 2026.

How many students use AI?

In the HEPI/Kortext 2026 survey of 1,054 UK full-time undergraduates, 95% used AI in some form and 94% used generative AI to help with assessed work. In 2024, 66% reported any use.

Are AI detectors accurate?

It depends on the tool, the text, and the error type. An independent 2025 benchmark found three commercial detectors kept false positives under 1%, with Pangram lowest. Accuracy drops on very short passages and on edited or “humanized” text, and vendors’ own figures often use different definitions of error.

Do AI detectors discriminate against non-native English speakers?

A 2023 Stanford study found seven detectors misclassified 61.2% of TOEFL essays as AI-generated. Newer detectors report better results, but few publish figures for non-native writers that independent researchers have verified. Caution is still warranted.

Can teachers prove a student used AI?

Not from a detector score alone. Turnitin’s own guidance describes the score as an indicator, not proof. Human markers are also unreliable here: Reading’s blind test found 94% of unedited AI exam answers went undetected.

What is the best alternative to AI detection?

There is no single replacement. The approaches with the most support are process-based assessment (drafts, oral follow-ups), a mix of supervised and open-AI assessments, and assignments tied to class-specific material.

Sources

Primary and peer-reviewed

• Liang, W., et al. (2023). GPT detectors are biased against non-native English writers. Patterns. arXiv version · Stanford HAI summary
• Scarfe, P., et al. (2024). Blind test of AI-generated exam answers at the University of Reading. PLOS ONE. University of Reading release
• Jabarian, B., & Imas, A. (2025). Artificial Writing and Automated Detection. NBER Working Paper 34223. PDF · Chicago Booth Review summary
• Stephenson, R., & Armstrong, C. (2026). Student Generative AI Survey 2026 (HEPI Report 199, sponsored by Kortext). HEPI

Journalism and secondary reporting

• The Guardian FOI investigation of 131 UK universities, June 2025 (add article link before publishing).
• CNN, reporting on Stanford’s high school survey: syndicated copy
• The Decoder, coverage of HEPI 2026: article
• Higher Ed Dive, on Turnitin’s false-positive disclosures: article
• The Wall Street Journal, blue-book sales, May 2025 (add article link before publishing).

Vendor material (treat as vendor-reported)

• Turnitin, Chief Product Officer update on AI writing detection: blog
• GPTZero, re-evaluation on the Chicago Booth benchmark, 2026: GPTZero

Methodology note: figures were checked against the primary source or the original publisher where available. Vendor-reported numbers are labeled as such. Where an independent source could not be found, the article says so rather than filling the gap. This article reflects evidence available as of October 21, 2026; detector products and institutional policies change quickly, so check current vendor documentation before making decisions about individual cases.