What OpenAI announced
The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way. OpenAI's answer, published on 5 October 2026, comes in three parts:
- ChatGPT and Codex in the EU. Over the coming weeks, eligible text output for users in the European Union, across all plans, gets an invisible watermark. OpenAI is deliberately not making it a global default.
- API customers, worldwide, opt-in. From the same day, API customers can switch on watermarking for select models. It stays off unless they do.
- A restricted detector. Applications are open, but access goes first to approved researchers and expert organisations, case by case. OpenAI cites the risk of missed watermarks and false positives as the reason it is not public.
| Content | Watermarked? | Who can check |
|---|---|---|
| ChatGPT & Codex text | Eligible output for users in the EU, rolling out over the coming weeks. Not a global default. | RestrictedOnly by approved researchers and expert organisations, case by case. |
| API text | Opt-in for select models, worldwide. Off by default. | RestrictedSame restricted detector. |
| Images | C2PA content credentials plus an invisible SynthID watermark on supported outputs. | PublicPublicly, through openai.com/verify and the Content Provenance API. |
| Audio | Invisible SynthID watermark on supported outputs. | PublicPublicly, through openai.com/verify and the Content Provenance API. |
The background is Article 50 of the EU AI Act, whose transparency rules took effect on 2 August 2026. Anthropic answered the same rule by watermarking Claude output worldwide; OpenAI chose to start in the EU and learn from real-world use first.
How textGrain works
textGrain adds an invisible statistical signal to the model's word choices. The reader sees normal text; the detector tests whether a passage carries the pattern. It is the same family of idea as Google's SynthID, and OpenAI reports that textGrain matched or exceeded the other approaches it tested, SynthID for text included. It has published a technical report and says it plans to release the technology as open source.
Output quality held up in OpenAI's tests: across eight benchmarks for its frontier model it reports no meaningful difference with and without the watermark — GPQA Diamond, for example, at 94.44% without and 93.94% with, and some scores slightly higher with it.
How well it works, by OpenAI's numbers
OpenAI is unusually direct about the limits. All figures below are at a target false-positive rate of 1%:
- Length matters. About 80% of 200-token passages were detected, against about 95% of 400-token passages, on topics such as psychology.
- Topic matters. Detection was substantially lower for content like mathematics, where there is less freedom in word choice.
- Editing matters most. On 400-token passages, replacing 10% of the words with synonyms dropped detection from about 92% to 66%. Replacing 25% dropped it to 17%.
A 1% false-positive rate also means roughly one in a hundred unwatermarked passages can be flagged. That is a reasonable research setting, and one reason OpenAI is not putting the detector in front of teachers or employers yet.
What a watermark can't tell you
OpenAI lists the limits itself, and they are worth repeating for anyone tempted to treat a result as proof:
- It does not measure human contribution — only that an OpenAI system generated or processed part of the passage.
- It does not establish ownership, responsibility, or whether disclosure was required.
- It does not identify the user, account, prompt or conversation.
- It does not say whether the text is accurate.
- A missing watermark does not prove a human wrote it: the text may be short, edited, translated, from an unwatermarked model, from before the rollout, or from another company's tools.
What this means if you check ChatGPT text
For now, very little changes in practice. Most ChatGPT text in circulation was written before the rollout, outside the EU, or through the API with watermarking off — and the detector is not open to schools, publishers or employers. Checking whether text was written by ChatGPT still means a classifier that reads the writing itself, not a key.
That is what tropa-3, our text model, does: it scores text from ChatGPT, GPT-6, Claude, Gemini and other models without relying on any watermark, so it works on text from any provider, any region and any date. The September 2026 benchmark has its numbers next to GPTZero and Pangram. Like a watermark, a score is evidence, not proof.
The two approaches complement each other. A watermark, where present, is strong evidence that OpenAI was involved; a classifier covers everything the watermark never reaches.
Images and audio: checkable today
The restrictions apply to text only. OpenAI already adds C2PA content credentials and a SynthID watermark to supported images and a SynthID watermark to supported audio, and anyone can verify them through openai.com/verify or the Content Provenance API.
WasItAIGenerated runs that OpenAI check automatically. Every image you upload is checked by our own detector and, in parallel, by OpenAI's provenance API; if OpenAI confirms its credentials or watermark, the result is marked as verified. The audio check looks for OpenAI's voice watermark. Both work on the website, and images also through our API and MCP server.
Check text, images or audio for AI
Text runs on tropa-3. Images and audio include OpenAI's provenance check.
Try it freeFAQ
Does ChatGPT watermark its text?
Starting in October 2026, yes — but only partly. OpenAI announced on 5 October 2026 that it will add an invisible watermark called textGrain to eligible ChatGPT and Codex text for users in the European Union over the following weeks. Outside the EU it is not a default. API customers anywhere can opt in for select models; it is off unless they do.
How does the OpenAI text watermark work?
textGrain adds an invisible statistical signal to the model’s word choices. The text reads the same, but OpenAI’s detector can test a passage for that pattern. OpenAI says it matched or exceeded the other approaches it tested, including SynthID for text, and plans to release the technology as open source.
Can I check whether a text has an OpenAI watermark?
Not yet, for most people. OpenAI opened applications for its text watermark detector, but access is initially limited to approved researchers and expert organisations, granted case by case. It is not publicly available at launch. To assess ChatGPT text today you still need a classifier-based AI detector.
How reliable is the ChatGPT text watermark?
It depends on length, topic and editing. At a 1% false-positive rate, OpenAI’s detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages on topics such as psychology, and substantially fewer on mathematics. Replacing 10% of the words with synonyms cut detection of 400-token passages from about 92% to 66%; replacing 25% cut it to 17%.
If no watermark is found, did a human write it?
No. OpenAI states that a missing watermark does not prove human authorship. The text may be too short, edited or translated, come from a model that is not watermarked, predate the rollout, come from outside the EU, or come from another company’s tools.
Does a detected watermark prove someone cheated?
No. OpenAI says a watermark can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgement, editing or creativity went into it. It does not identify the user, establish ownership or responsibility, or say whether the text is accurate.
Can I check ChatGPT images and audio?
Yes. OpenAI adds C2PA content credentials and a SynthID watermark to supported images and a SynthID watermark to supported audio, and lets anyone verify them through openai.com/verify and its Content Provenance API. WasItAIGenerated runs that check automatically alongside its own image detector, and its audio check looks for OpenAI’s watermark.
Why is OpenAI doing this now?
The EU AI Act requires providers of generative AI to make generated text identifiable in a machine-readable way. OpenAI calls its EU-only rollout a phased approach that reflects both those requirements and the current limits of text watermarking.