Why AI detectors are a poor test for writers

Last updated 22 October 2026

An AI detector reads a piece of text and estimates how likely it is that a model wrote it. It is an estimate, and the evidence says it is often wrong.

Even the people who build the models struggled

OpenAI released an AI text classifier in January 2023. On its own challenge set, it correctly flagged 26% of AI-written text and wrongly labelled human-written text as AI 9% of the time. On 20 July 2023, OpenAI withdrew it “due to its low rate of accuracy” (OpenAI).

Detectors are harder on some writers than others

A 2023 Stanford study ran seven detectors on 91 TOEFL essays written by non-native English speakers. On average, the detectors labelled 61% of them as AI-generated, and 89 of the 91 essays were flagged by at least one detector. Essays by US eighth-graders were classified almost perfectly (Liang et al., 2023). Plain, careful writing looks “machine-like” to a detector, and that is the style many second-language writers use.

Why this matters if you use AI honestly

If you wrote a draft and used AI to tidy it, a detector may call it AI-written, human-written, or anything in between. None of those answers tells a reader what you actually did. And if you didn't use AI at all, a detector can still flag you.

A record beats a guess

Instead of asking a tool to guess, keep a record. Overwride stores the draft you wrote next to the rewrite it returned, and gives your post a footer that links to that Overwride record. A reader who wants to know what changed can open it and compare.

Overwride doesn't try to detect anything. Paste a post into the checker: if it exactly matches an Overwride rewrite, you see the record with its original draft. If not, the answer is always the same: we can't tell whether Overwride was used. That answer is less exciting than a score, and it never claims more than it knows.

Try it on your next post. Starter is €10 a month +VAT, and checking a post is free.