does chatgpt watermark your text now under the eu ai act
Not yet, and only in the EU when it does. OpenAI announced textGrain on October 5, 2026: an invisible watermark in ChatGPT and Codex text for EU users, rolling out “over the coming weeks” across every plan. API customers worldwide can switch it on today, off by default. Outside the EU, nothing changes.
is the chatgpt watermark live yet?
Three things shipped on three different schedules, which is where the confusion comes from.
The API toggle went live October 5. It’s off unless you turn it on, it’s available worldwide, and you choose which models get marked — organization-wide under Data controls → Text provenance, or overridden per project. Coverage reaches legacy models over the coming weeks.
ChatGPT and Codex in the EU: announced, not finished. The wording is “over the coming weeks,” on all plans, EU only. The date that actually binds is December 2, 2026. Article 50’s machine-readable marking rule took effect August 2, but systems already on the market before then got a grace period until December, and ChatGPT was very much on the market. So the real ship window is somewhere between now and then.
Everywhere else: nothing. OpenAI said it is not making this a global default at launch. That’s the opposite call from Anthropic, which switched marking on worldwide in August rather than run one pipeline for Europe and a clean one for everyone else.
how does openai’s text watermark work?
At each step the model holds a ranked set of probabilities for the next token. textGrain builds several adjusted versions of that set, each favoring different candidates, balanced so that averaging them returns the model’s original distribution. A secret key selects which version gets used, and the model samples from it. Over a few hundred tokens those choices lean in a direction a detector holding the same key can measure.
What it isn’t: hidden characters. No zero-width spaces, no invisible Unicode, no tell-tale punctuation. OpenAI states outright that copying and pasting introduces no hidden material — which kills the entire cottage industry of watermark removers that scrub invisible bytes, the same wrong target they already aim at with Claude.
The accompanying technical report is 20 pages of real math — optimal transport, Gumbel costs, KL regularization — and contains not a single detection percentage. Every number below comes from the blog post and the help center instead. The report is promised an update “in the coming weeks.”
how accurate is chatgpt watermark detection?
At a target false positive rate of 1%, on English passages about something discursive like psychology: roughly 80% of 200-token passages, roughly 95% at 400 tokens. Math questions land substantially lower. When there’s one correct way to write the answer, there’s no slack left to encode anything in.
Editing is the real ceiling. On 400-token passages, swapping 10% of the words for synonyms dropped detection from about 92% to 66%. Swapping 25% took it to 17%. That’s one word in four — not a rewrite, not a paraphrasing tool, just light sanding.
does it work in the other eu languages?
Worse, and nobody put this in a headline. OpenAI ran all 24 official EU languages: 500 synthetic English prompts, translated into the other 23. At the same 1% false positive rate, Spanish detected best at 69.0% and Romanian worst at 42.2%.
So a watermark built for the EU performs in EU languages somewhere between a coin flip and two in three. The 95% figure is English. There’s an adjustable strength knob and OpenAI says it turned that up for languages scoring under 60%, which helps — and also means speakers of those languages get a heavier mark than the one that got benchmarked.
does it watermark short answers and code?
Mostly no, and that’s sanctioned. The EU Code of Practice on transparency doesn’t require marking outputs shorter than 200 tokens — about 150 words of English — or code snippets. Short replies don’t offer enough token choices to detect, and code has too few plausible alternatives at each step to hide a signal in.
Which makes Codex a strange thing to put in the announcement. The product that writes code is the one whose output is hardest to mark and exempt anyway.
can you check if text is watermarked?
No. Not as a teacher, an editor, a publisher, or a developer paying OpenAI for API access.
The detector exists and applications opened October 5, but access is granted case-by-case to approved researchers and academic organizations, and OpenAI says plainly it isn’t public at launch, citing missed watermarks and false positives. Turning on watermarking for your own traffic doesn’t get you in — the help center answers that one directly. Meanwhile openai.com/verify stays open to anyone, for images and audio. Text is the single medium where the signal exists and the verification doesn’t.
Third-party detectors like Pangram are a different animal: classifiers guessing from style after the fact, not readers of an embedded mark. The AI Act asks for the embedded mark. Only OpenAI can read OpenAI’s.
can you turn the chatgpt watermark off?
In the API, it’s off until you switch it on. For EU ChatGPT users, OpenAI hasn’t said whether an opt-out exists, and I’m not going to invent one — there’s nothing in the announcement, the help center or the FAQ about disabling it. Treat it as mandatory until they say otherwise.
does watermarking make the output worse?
No measurable difference. Across the benchmark suite for Astra, OpenAI’s current frontier model, watermarked and unwatermarked scores sit inside normal run-to-run variance and trade places in both directions: GPQA Diamond 93.94 against 94.44, Terminal-Bench 4.0 56.06 against 53.90. Speed impact is described as negligible, and earlier ChatGPT tests showed no movement in thumbs-down rates.
Article 50 asked for one thing: generated text that machines can identify as generated. What shipped is a signal one company can read, that fades to 42% in Romanian, dies when you change one word in four, skips anything under 150 words, and can’t be checked by the people the transparency was written for.
Compliance: achieved. Transparency: pending.