OpenAI’s text watermark is a useful signal and a poor verdict. It works best on long, unedited passages and fades quickly when text is shortened or reworded. Even at a 1% false-positive rate, it would wrongly flag about one in a hundred genuinely human texts. If you’re a teacher, editor or publisher, use a positive result to start a conversation, never to end one. And if you’re outside the EU, ordinary ChatGPT answers aren’t being watermarked at all, despite what some headlines suggest.
I’ve used AI text detectors myself, and their results seemed acceptable to me. Still, a detector’s guess and a watermark are different things, and I don’t think either should be the only evidence used to judge someone’s writing.
Does ChatGPT Watermark Text? The Short Answer
| Where | Is text watermarked? |
|---|---|
| ChatGPT and Codex in the European Union | Yes, rolling out over the coming weeks for eligible users across plans |
| ChatGPT outside the EU | No, not by default |
| OpenAI API, worldwide | Only if you opt in, for select models (off by default) |
OpenAI announced this on 5 October 2026 in response to the EU AI Act. The technology is called textGrain.
How textGrain Works
A language model picks each next word (token) from a set of possibilities. OpenAI says textGrain nudges those choices according to a secret pattern, so a compatible detector can later find a statistical signal across a long enough passage. The mark lives in the wording itself, not in hidden characters, odd spacing or metadata, so copying and pasting doesn’t remove it. Editing, shortening, translating or mixing with human writing can.
OpenAI says its evaluations showed no meaningful quality difference between watermarked and normal output, and a negligible effect on speed. Those are OpenAI’s own measurements.
How Reliable Is Detection?
OpenAI is unusually open about the limits. At a 1% false-positive target, its detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages in one evaluation built on ELI5 questions. Detection was much weaker for constrained content like maths, where there are fewer ways to phrase things. Editing hurt badly: in a 400-token test, swapping 10% of words for synonyms cut detection from roughly 92% to 66%, and swapping 25% cut it to 17%. OpenAI says textGrain matched or beat other methods it tested, including Google’s SynthID for text, but 200 tokens is the shortest passage in its results.
That’s why there’s no public “AI checker”. Detector access is limited to approved researchers and expert organisations, case by case, and the tool only reports whether it found an OpenAI watermark. It doesn’t reveal user identities, prompts or chats.
Our Simulation: How Length and Editing Affect a Statistical Watermark
Method: OpenAI hasn’t published textGrain’s design, and its detector isn’t public, so we couldn’t test textGrain itself. Instead, on 15 October 2026 we simulated the best-known published design of this kind, the “green list” watermark described by Kirchenbauer and colleagues in 2023. In that scheme, a secret key marks a quarter of possible next words as “green” at each step, the model is nudged toward them, and a detector counts how many green words a passage contains. We ran 20,000 simulated passages per cell in Python (with AI assistance), set the detector to a 1% false-positive rate, and modelled each edited word as also scrambling the next word’s check, because the green list depends on the previous word. This is a model of the general technique, not a measurement of ChatGPT.
| Passage length (tokens) | No edits | 10% of words changed | 25% changed | 40% changed |
|---|---|---|---|---|
| 100 | 82% | 62% | 26% | 5% |
| 200 | 99% | 92% | 52% | 9% |
| 400 | 100% | 100% | 84% | 17% |
| 800 | 100% | 100% | 99% | 36% |
The table shows the share of watermarked passages detected for varied prose, where the model has plenty of word choices. When we lowered the signal to mimic constrained text such as maths, detection of unedited 400-token passages fell from 100% to 79%, and to 24% with a quarter of the words changed.
What I take from it
- Length is the biggest lever. Going from 100 to 200 tokens took detection from 82% to 99% in our model. A short post, or one AI paragraph inside a human essay, gives a detector very little to work with.
- Real edits seem to hurt more than our model predicts. OpenAI’s own figures fall faster than ours (92% to 66% at 10% synonym swaps; 17% at 25%), which suggests real rewording is harder on textGrain than random edits. That strengthens the case against relying on it.
- A 1% error rate is still a lot of people. Our simulated detector flagged 1.0% of unwatermarked text, as designed. A university that checks 10,000 genuinely human essays a term would still wrongly flag about 100. OpenAI itself wrote in 2024 that low false-positive rates still produce many false positives at scale.
What a Positive Result Does and Doesn’t Prove
A detected watermark is evidence that an OpenAI system generated or processed some of a passage. It doesn’t identify who used the model, reveal the prompt, say how much a human contributed, or prove the content is accurate. And a missing watermark proves nothing: the text may be short, edited, translated, made before the rollout, made by an unsupported model, or made by another company’s AI.
My advice for publishers and schools: keep drafts, revision history and sources for important work, and never treat a detector result as proof of authorship on its own.
Why the EU, and Why Now
This section draws on Search Engine Journal’s report, which I read in full. Article 50 of the EU AI Act began applying on 2 August 2026, and systems already on the market have until 2 December to meet its marking and detection obligation. A voluntary Code of Practice on AI-generated content offers a way to show compliance; about 190 organisations had signed by the end of July, and OpenAI publicly supported it in June. It’s a reversal for OpenAI, which shelved ChatGPT text watermarking in 2024 after a survey found almost 30% of users said they’d use ChatGPT less if it were watermarked.
One practical consequence the report notes: a team split between Europe and elsewhere, sharing one ChatGPT plan, could get watermarked text in some offices and not others.
How It Compares with Claude
Summarised from Search Engine Journal’s comparison, which cites Anthropic’s own pages; I haven’t checked those pages myself.
| OpenAI | Anthropic | |
|---|---|---|
| Apps | Eligible ChatGPT and Codex text, EU only | Supported Claude models, worldwide |
| API | Opt-in for unnamed select models, off by default | Marked on supported models, listed in a support article |
| Method | textGrain, OpenAI’s own | A version of Google DeepMind’s SynthID-Text |
| Detector access | Approved researchers and expert organisations | Private preview for eligible organisations, including regulators, media and researchers |
| Editing results published | Yes (10% and 25% synonym swaps) | No figures in its explainer |
The report says Anthropic marks text worldwide because it doesn’t yet have a durable way to limit watermarking by region. Claude competes with ChatGPT, and Claude is also made by the company behind the assistant I use to help write this site, so check both companies’ pages for the current position.
What’s Still Unknown
- Which API models are included in the opt-in.
- Whether EU ChatGPT users can turn the watermark off.
- When OpenAI will widen detector access.
- How well the mark survives everyday translation.
OpenAI also says it’s working with cloud partners such as Microsoft Azure to offer the API watermark through their services.
FAQ
Can I remove the ChatGPT watermark?
Heavy editing weakens it a lot, as both OpenAI’s figures and our simulation show. But if you’re relying on AI text where disclosure is expected, the honest answer is to disclose it.
Can I check whether text has a watermark?
Not yet as a member of the public. Detector access is limited to approved researchers and organisations.
Does the watermark affect quality?
OpenAI says its tests showed no meaningful difference.
Related: ChatGPT vs Claude vs Gemini · AI agent safety
More from EvalAITools: OpenAI Dots explained · Gemini connected apps and privacy
Sources (checked 13 to 15 October 2026)
- OpenAI: Our approach to EU text provenance rules
- OpenAI Help Center: Provenance signals in OpenAI-generated content
- Search Engine Journal: OpenAI to watermark ChatGPT text in the EU (5 October 2026)
- Kirchenbauer et al., “A Watermark for Large Language Models” (2023), the scheme used in our simulation
