PPROMPT WORKBOOK Build a prompt
Verified 27 July 2026

AI detectors and false accusations

Almost everything written about AI detection explains how to avoid it. This page is about the other side: the false positives, who they land on, and how to respond.

The company that built the model gave up on detecting it

OpenAI launched an AI Text Classifier in January 2023 and discontinued it on 20 July 2023, less than six months later, citing a low rate of accuracy. By OpenAI's own disclosure at launch, the tool correctly identified AI-written text only about 26 percent of the time, and mislabelled human writing as AI-generated roughly 9 percent of the time.

That is the single most useful fact in this whole debate. The organisation with the deepest possible knowledge of how its own model writes could not build a reliable detector for it, and said so publicly.

The bias that makes this an equity problem

A peer-reviewed Stanford study published in Patterns in 2023 tested seven GPT detectors against essays written by non-native English speakers. It reported that the detectors misclassified genuine TOEFL essays as AI-generated at an average rate of around 61 percent, while classifying essays by US school students accurately. Nearly all of the TOEFL essays were flagged by at least one detector.

The mechanism is not mysterious and it is the uncomfortable part. Detectors largely key on linguistic simplicity and predictability. Writing produced by someone still building fluency in English tends to use a narrower vocabulary and more regular constructions, which is exactly the signature these tools read as machine-generated.

A note on our own sourcing. The study is peer-reviewed and we link it below. The specific percentages above are consistently reported across sources, but the publisher blocked automated access to the full text, so we could not read the figures off the original page ourselves. Treat them as reported rather than as something we personally verified line by line.

What the vendors admit

Turnitin launched its detector in April 2023 claiming a false-positive rate below one percent. By June 2023 its own chief product officer publicly acknowledged a real-world sentence-level false-positive rate of around four percent. GPTZero publishes a figure of no more than one percent, which remains vendor self-reported; we found no independent audit confirming it.

Four percent sounds small until you multiply it. Across a class of thirty students writing weekly, sentence-level flagging at that rate produces false accusations continuously rather than occasionally.

If you are accused and you did not do it

This is stressful and the burden of proof tends to land on the accused, which is backwards. A survey of students commissioned in early 2026 found that 75 percent of those using AI reported significant stress about being wrongly flagged. It is worth noting that survey was commissioned by a company in the sector rather than conducted independently, but the direction is unsurprising.

  1. Ask what the actual evidence is

    A detector score is a probability estimate from a tool with a known error rate, not evidence of misconduct. Ask which tool was used, what score it produced, and what that institution's own policy says the score means.

  2. Produce your process, not your innocence

    You cannot prove a negative. You can show version history, drafts, notes, search history, or a document's revision timeline. This is why keeping drafts matters even when nobody is asking.

  3. Point to the vendor's own numbers

    Turnitin's four percent acknowledgement and OpenAI's withdrawal of its own classifier are both on the record and both cited below. They are stronger than any argument you make about yourself.

  4. Raise the bias evidence if it applies to you

    If English is not your first language, the peer-reviewed finding that detectors misclassify non-native writing at high rates is directly relevant to your case.

  5. Offer a live demonstration

    Writing a comparable passage under observation, or talking through your reasoning and sources, is usually more persuasive than any document.

What not to do. Do not run your work through a “humaniser” tool to lower a detector score. If you wrote it yourself you are now disguising your own writing, which looks far worse if it comes out, and it concedes the premise that the score was meaningful.

If you are the one setting policy

The evidence does not support using detector scores as the basis for an accusation. It can support using them as one weak signal that prompts a conversation, provided the conversation is genuinely open. A policy worth having states which tool is used, what threshold triggers what, that a score alone is never sufficient, and how someone appeals. A policy that leaves all of that unstated puts the entire cost of the tool's error rate on the least powerful person in the room.

Sources