Do AI detectors work?
Short answer: not well enough to accuse anyone. Here is why, and what works better.
What the evidence says
- OpenAI withdrew its own AI text classifier in 2023, citing its low rate of accuracy.
- False positives hit some people more. A 2023 Stanford study (Liang and colleagues) found popular detectors flagged more than half of essays written by non-native English speakers as AI-generated.
- Light editing defeats them. Rewording, adding typos or asking the chatbot to "write casually" often changes the verdict.
Why they struggle
Detectors look for text that is statistically predictable. But careful human writing can be predictable too (formal essays, simple English, templates), and AI can be prompted to be less predictable. The two overlap too much for a clean line.
What to use instead
- Ask about it. A person who wrote something can talk about it, explain a choice or add a detail.
- Look at the process: drafts, notes, version history.
- Read for habits, not proof: the signs of AI writing are a reason to look closer, never a verdict.
In conversation, the best detector is a good question. The daily game trains exactly that.
Questions
Is the AI-o-meter on this site a detector?
No. It is a toy that counts writing habits, and says so. Don't use it, or any detector, as proof.
My work was flagged as AI but I wrote it. What can I do?
Show your drafts or version history, offer to talk through how you wrote it, and point out that detectors make documented mistakes.