Why AI detectors aren’t the answer — and what to do instead

When students started using AI to write, AI-writing detectors seemed like the obvious fix. The evidence since then has been sobering, and many schools and universities have stepped back from them.

The problem with detectors

They make false accusations. Detectors estimate how “AI-like” a piece of writing is. Clear, simple, predictable prose can look AI-like even when a student wrote every word.

They’re biased against multilingual writers. A widely cited Stanford study tested seven popular detectors on 91 essays written by non-native English speakers for the TOEFL exam. On average, the detectors labeled 61% of these human-written essays as AI-generated, and 97.8% were flagged by at least one detector. The same detectors classified essays by U.S. eighth graders accurately — the difference was the writers, not the honesty of the work. Students who write in a second language often use simpler vocabulary and sentence structures — exactly what detectors are tuned to notice.

They’re easy to fool. Light rewording or running text through a paraphrasing tool often changes a detector’s verdict, so the students most likely to be caught are the ones who didn’t try to hide anything.

Even AI companies backed away. OpenAI shut down its own AI-text classifier in 2023, citing its low rate of accuracy.

A false accusation can damage a student’s trust in you, and their relationship with school, for a long time. A detector score on its own isn’t evidence.

What to do instead

Make the rules explicit

Students are getting mixed messages: in RAND’s survey data, only about a third said their school had a schoolwide AI policy. For each assignment, say clearly whether AI is not allowed, allowed for specific steps (brainstorming, checking spelling), or welcome throughout — and ask students to note how they used it. A simple traffic-light label (red, yellow, green) on the assignment works well.

Design for process, not just product

• Write in class for key pieces, or at least start them in class.
• Collect drafts and notes along the way, not only the final version.
• Hold short conferences. Two minutes of “walk me through your main point” tells you more than any detector.
• Ask for reflection. “What was the hardest part to write, and why?”
• Connect to what only your class knows — a class discussion, a local place, a shared experience, the student’s own life.

Assess thinking in more than one way

Oral explanations, diagrams, annotated sources, and “show it another way” tasks make it much harder for AI to stand in for understanding — and they’re good teaching anyway. Our post on exit tickets has quick formats.

If you suspect misuse, start with a conversation

Ask the student to talk through their work, explain a choice they made, or continue the piece in front of you. Approach it with curiosity, not accusation, and follow your school’s academic integrity process. Most of the time the conversation itself is the most useful evidence you’ll get — and the most useful teaching moment.

Teach students how to use AI well

Students who understand how AI works, where it goes wrong, and when it helps their learning are better equipped to make good choices. See teaching AI literacy.

Sources

• Liang et al., GPT detectors are biased against non-native English writers (Stanford; published in Patterns, 2023)
• OpenAI, New AI classifier for indicating AI-written text (updated July 2023 when the classifier was discontinued)
• eSchool News, Student use of AI for homework rises (reporting RAND data, March 2026)