Atlas AI Resources

← Back to AI News

AI policy

ChatGPT Text in Europe Will Carry a Hidden Watermark. It Proves Less Than You Think.

Published October 6, 2026

On Monday, October 5, OpenAI said it will add an invisible watermark to text that ChatGPT and Codex, its coding assistant, write for users in the European Union. The goal is to meet the EU AI Act’s rule that AI-generated text be identifiable by software. The mark is hidden in the word choices themselves, nobody outside a small research group can check for it yet, and OpenAI’s own tests show that light editing weakens it a lot.

What happened

A watermark here is not a logo or a hidden character. It is a faint statistical pattern in which words the model picks. OpenAI calls its method textGrain. As a model writes, it scores every possible next token (a word or piece of a word). textGrain builds several slightly different versions of those scores, each nudged toward different words by a secret key, and balances them so that on average they match the original. A detector holding the same key can later check whether a passage follows that pattern more often than chance would allow. TechCrunch reports the mark survives copy and paste, and OpenAI’s help page says it does not insert hidden characters.

The rollout has three parts. Over the coming weeks, eligible ChatGPT and Codex users on every plan get watermarked text, but only in the EU. Starting Monday, developers anywhere who use OpenAI’s API (the paid connection businesses use to build OpenAI models into their own apps) can switch watermarking on for some models; it is off by default. And approved researchers and expert organizations can apply to use the detector. OpenAI says it is not making the detector public because of the risk of missed watermarks and false positives (flagging text that carries no watermark).

TechCrunch reports that the AI Act’s transparency rules took effect on August 2, and that Anthropic said two months ago it would watermark text from its Claude models worldwide.

Why it matters to a normal person

If you live in the EU, text you get from ChatGPT will soon carry a mark you cannot see. OpenAI says the mark does not identify you, your account, or your prompt, and that its tests found no meaningful drop in answer quality. Those are the company’s own findings.

The limits matter more than the headline. In OpenAI’s tests, with the detector tuned to wrongly flag about 1% of unmarked text, it caught about 80% of 200-token passages and about 95% of 400-token passages on topics like psychology. Math answers, where there are fewer ways to say the same thing, were caught far less often. Swapping 10% of the words in a 400-token passage for synonyms cut detection from about 92% to 66%, and swapping 25% cut it to 17%. The EU’s code of practice on AI-generated content does not require watermarks on “very short text,” which it defines for now as text shorter than 200 tokens.

What it does not mean

It does not mean teachers or employers can now run an essay through an OpenAI checker. They cannot, at least not yet. Outside AI-detection tools work differently: OpenAI’s help page says they typically judge finished text by patterns such as word choice, rather than looking for a signal planted during writing.

It does not prove much even when it fires. OpenAI says a watermark can show that one of its systems generated or processed part of a passage, but not how much a person contributed, who owns the text, or whether it is true. A detector result is a clue, not a verdict.

And no watermark does not mean a human wrote it. The text could be short, edited, translated, older than the watermark, or written by another company’s AI. OpenAI also says it is not turning this on for ChatGPT outside the EU at launch.

Sources