Does ChatGPT Watermark Its Output?
Chapter 2, part 1 of 3 · 6 min read — by HALLMARK.AI
Mostly yes, with gaps. OpenAI says images made in ChatGPT, Codex and its API carry both C2PA metadata and a SynthID watermark. For text, OpenAI announced in October 2026 that it would watermark ChatGPT and Codex text for users in the EU, with an opt-in for API developers. Checking is limited: a preview image checker exists, and the text detector is not public.
Images: two signals
- C2PA metadata. Content Credentials are a signed record attached to the file. OpenAI has attached them to ChatGPT images since 2024. They are hard to forge but easy to delete.
- SynthID. An invisible watermark from Google DeepMind, added in May 2026 according to OpenAI's help centre and news coverage of the announcement. Because it sits in the pixels, it can survive some edits that remove metadata. How SynthID holds up.
A visible label is optional: OpenAI's help article says you can request one by prompting the model to include it. It is not added by default.
Text
OpenAI shelved text watermarking in 2024. That changed on October 5, 2026, when it announced a method it calls textGrain, as reported by BleepingComputer from OpenAI's announcement. It shifts word choices to leave a statistical pattern, in the same family of idea as the marks Claude puts in text. Per that coverage:
- ChatGPT and Codex text is watermarked for EU users on all plans, rolling out over the weeks after the announcement. It is not a global default.
- API developers anywhere can opt in for supported models. It is off by default.
- OpenAI's own tests show light editing weakens it: replacing 10% of words with synonyms cut detection in 400-token passages from about 92% to 66%, and 25% cut it to 17%.
- A detected mark does not reveal the user, account or prompt, and no mark does not prove a human wrote the text.
We could not confirm from OpenAI's own pages that the rollout is complete as of November 14, 2026.
Video and audio
OpenAI's provenance pages have also covered Sora video and audio. Reports differ on which signals Sora files carry, and we could not confirm them from OpenAI's own pages, so we make no claim here.
Can you check?
For images, partly. OpenAI previews a checker at openai.com/verify: you upload one image and it reports whether it finds C2PA metadata, a SynthID watermark, or no supported signal. At launch it covers OpenAI-generated content only. The text detector is not public. OpenAI says access is limited to approved researchers and expert organisations for now.
The limits
- Metadata strips. Screenshots, re-encoding and many upload paths drop C2PA. Its absence proves nothing.
- One tool only. These marks say a model may have generated a file. They do not say who made it, and they say nothing about a photo you took.
- Edits weaken text marks. As the numbers above show.
Where Hallmark fits
OpenAI's marks identify the tool. Hallmark's invisible watermark identifies the account that registered a file, and gives it a public certificate with a first-seen date. Our checker at /check is free and public. It does not tell you whether AI made a file. If you generate in ChatGPT, you can add your own mark to the final result with the ChatGPT connector.