AI Text Detectors: Why You Cannot Rely on Them
By Anis Hammouche·October 25, 2026
No, AI text detectors are not reliable. OpenAI withdrew its own in July 2023 because of its low accuracy, and a 2023 study found that these tools often classify as AI-written the texts of people whose native language is not English. If you write with AI, the risk concerns your reputation and your relationship with the client, far more than a technical test.
What OpenAI said when it stopped its tool
In 2023, OpenAI published a classifier meant to indicate whether a text had been written by an AI. In a note added to the launch announcement, the company says the tool has been unavailable since 20 July 2023 because of its low accuracy rate. At launch, it already acknowledged that the tool wrongly labelled about 9% of human-written texts as AI-written. Search Engine Land reports these details.
If a tool built by the company that makes the models did not hold up, you should doubt tools that present themselves as universal detectors.
The problem with false positives
A false positive is a human-written text classified as written by an AI. A 2023 study by Liang and co-authors, titled "GPT detectors are biased against non-native English writers" (preprint on arXiv), tested several detectors. Its authors conclude that these tools reliably identify texts by native speakers, but wrongly classify texts by non-native writers as AI-generated. They also warn against using them in evaluative or educational settings.
For you, the lesson is simple: a detector score does not replace your judgement on a text you sign. It can, however, flag a passage that is too generic, which you will then rewrite with a precise example from your work.
A detector therefore cannot establish with certainty that a text was generated by an AI. It must never be used as proof against someone.
What a score measures, and what it does not
A score is an estimate computed from the text alone. It does not know who wrote the text, how, or with which tool. It cannot see your notes, your drafts or your exchanges with the client. It therefore answers a different question from the one the accusation asks: "does this text resemble the ones the tool has seen?", not "who wrote this text?".
A very smooth text, with no example and no detail, can get an "AI" score even if it was written by hand. That is why the thing to fix is not the rhythm of your sentences, but the lack of concrete material.
What it changes for you
You may find yourself in three situations. The table sums up what each one implies.
| Situation | What the detector says | What you do |
|---|---|---|
| A client or an editor has your text analysed | An "AI" score is possible even if you wrote everything yourself | Show your process: outline, notes, sources, successive versions |
| You write in French with an assistance tool | The studies cited are about English, they do not measure French | Do not rely on any score, check the substance |
| You want to prove you did not use AI | A detector proves nothing, in either direction | Keep your drafts and sources |
The most concrete point concerns proof of work. If doubt arises about an important text, your history is worth more than any score: research notes, time-stamped drafts, sources consulted, exchanges with the client. Keep them while you produce the text, not after the fact.
If a detector challenges your text
When a score lands on your desk, stay calm and proceed in this order.
- Ask which tool was used, which version and which threshold. A score without a method proves nothing, and two tools can give different results on the same text.
- Show your process. Present the outline, the research notes, the sources you opened and the successive versions. That file carries more weight than a percentage.
- Discuss the substance. Ask which passages are the problem and why. Someone who knows their subject explains their choices in detail. A text reworked by a tool does not hold up well under that kind of discussion.
- For a high-stakes case, read the written rule. A competition, a school, a contract or a publication has its own guidelines. If they are unclear, ask for a clarification in writing.
What to keep in mind when you review
Your text must be accurate, answer your reader's question and sound like you. A text that is accurate and precise has nothing to hide, whichever detector reads it.
When you review, look for three things:
- factual claims, each with its source opened and dated;
- generic sentences that could appear in any article on the topic, and that you can replace with a precise example from your work;
- hollow phrasing, such as "a major challenge" or "an innovative approach", which says a lot without claiming anything.
This review improves the text whatever its origin. It is equally useful for a handwritten text.
The risk exists even if you only use AI to correct a typo or rephrase a sentence. A text modified by a tool is still your text: you answer for its claims, not the detector or the tool.
Frequently asked questions
Are AI text detectors useful at all? They give an indicative signal, never proof. A teacher or a client can use one as a starting point for a question, without treating the result as a verdict.
Can someone be wrongly accused? Yes, that is the risk the study cited describes: texts by non-native writers were classified as generated. A person writing in a language they learned is therefore particularly exposed.
Should you say that you used AI? It depends on the context. Read the rules for your situation before publishing, and ask for a clarification in writing if they are unclear.
Must a text sound human to be good? No. It must be accurate, clear and useful. The rest, fluency or personality, comes from careful review and real examples.
Stuck on a specific case? Write to me through the contact page.
