You Can't Detect Your Way Out of the AI Cheating Problem

By 2026, roughly 92% of college students report using AI tools for coursework — up from 43% in 2023. Sixty percent of higher ed leaders say cheating has gone up since generative AI arrived. And by most estimates, the large majority of AI-written assignments are never flagged at all.

The institutional response has mostly been to buy a detector. That's the wrong move, and it's worth being clear about why.

Detection is an arms race, and the house always loses

AI-writing detectors don't work well enough to base decisions on, and the vendors say so themselves. Turnitin has publicly acknowledged that to keep its false-positive rate under 1%, it deliberately lets up to 15% of AI-written text through. Its own guidance warns that a score should never be the sole basis for accusing a student. Independent testing has produced false-positive rates far higher than the advertised number.

Even if the accuracy were perfect, the underlying game is unwinnable. Every improvement in detection is met with a paraphrasing tool, a "humanizer," a new model. You are committing your institution to permanent escalation against companies that ship faster than you can write policy. Plenty of universities have already figured this out and quietly walked back their detection programs after the dismissed-allegation rate got embarrassing.

The false positives land on the wrong students

Here's the part that should stop a provost cold. Detectors don't fail randomly. They flag non-native English speakers, first-generation students, and writing in technical subjects at higher rates — STEM and computer-science work especially, because clean human code and clean AI code look alike. In other words, the tool you bought to protect academic integrity is most likely to wrongly accuse the students who have the least margin to fight a misconduct charge.

A policy that produces that distribution of errors isn't a rigor problem. It's an equity problem wearing a rigor costume.

The number that actually matters

Buried in the cheating statistics is a more useful one: while AI use is nearly universal, only about 18% of students say they use it to do an assignment for them outright. Most students aren't trying to cheat. They're stuck at 11pm, the help desk is closed, and ChatGPT is right there and will give them the answer.

That's a design problem, not a character problem. And design problems have design solutions.

Build the tool so it won't do the work for them

If most "cheating" is really students reaching for the only help available, then the answer isn't catching them after the fact. It's changing what's available. A general-purpose chatbot is built to deliver the answer as fast as possible — that's the whole product. A teaching tool should be built to refuse.

That refusal is the entire point, and it's structural, not a personality you prompt into ChatGPT. A tutor built for learning should withhold the final answer and walk the student through the steps instead. It should recognize the difference between "explain why my approach is wrong" and "just give me the solution," and treat those requests differently. It should give the instructor signal about where the class is stuck — without putting individual transcripts on a surveillance dashboard. When the legitimate help students actually want is one tab over and it scaffolds instead of solving, the incentive to paste the prompt into a consumer chatbot quietly drops.

That's the version of academic integrity that scales. Not a detector that accuses your most vulnerable students of writing too cleanly. A tool that was never willing to do the assignment in the first place.

What to ask before you renew the detection contract

If detection is on your budget this year, three questions are worth asking first. What happens to a student we flag incorrectly — and who tends to get flagged? Are we spending to catch students, or to help them learn — and which one actually moves outcomes? Is there a version of this where the AI is on the instructor's side instead of the adversary's?

There is. We're building it for STEM, and we'd rather help students learn than catch them after they didn't. If your department is wrestling with this for fall, we'd love to talk — we're running pilots with universities and community colleges right now.