Who watermarks what
Of the 10 provider-and-modality combinations we track, 8 are confirmed shipping some form of marking, 2 have no way for anyone outside the company to verify the mark, and 2 are reported but unconfirmed because we could not find a primary source. Anthropic, Google and OpenAI have primary documentation of what they actually do; for Meta and Microsoft we have found none.
Updated 2026-10-23. Published as reusable data under CC BY 4.0. Corrections with a primary source are very welcome.
| Provider | Modality | Status | Method and scope | Can you verify it? |
|---|---|---|---|---|
| Anthropic Claude |
text | Partial | Statistical watermark in the model's word choices — a version of SynthID-Text, named by Anthropic on 14 Aug 2026[1][10][12] Models launched on or after 2 Aug 2026, worldwide. Since 16 Sep 2026 Anthropic's help centre lists which models carry the text watermark, and since 7 Oct 2026 it lists nine: Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5 (Opus 5.5 added after 21 Sep 2026, Sonnet 5.5 on 28 Sep 2026, Fable 5, Opus 4.8 and Sonnet 5 on 1 Oct 2026, Haiku 5.5 on 7 Oct 2026, the day Anthropic announced it). Seven other listed models (Mythos 5, Opus 4.5 to 4.7, Sonnet 4.5 and 4.6, Haiku 4.5) are listed for file credentials only. Anthropic says it is adding the watermark to models released before 2 Aug 2026; until 1 Oct 2026 it said all would be covered by 2 Dec 2026, and it now gives no date. Anthropic published a technical explainer on 14 Aug 2026 naming the scheme: a version of Google DeepMind's SynthID-Text, in the family of approaches going back to Scott Aaronson's 2022 proposal, which change only the source of the randomness used to pick among equally good next words. It also states plainly that nothing is added to the text and there are no hidden characters, and that the mark carries no information about a user, their organisation or their chats. This supersedes a note published here until 15 Aug 2026 saying Anthropic had not published the scheme's details. Marking is thinner where word choice is constrained: factual passages, proofreading, and code outside its comments. This row read 'None' in the detector column until 2 Sep 2026, which was correct until Anthropic opened the detection API to a private preview on 1 Sep 2026. |
Gated Since 1 Sep 2026 a detection API is in private preview. Anthropic offers it to organisations obliged to check under EU law — regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups — and to enterprises under the same obligation, by application through a form. It says access will widen over time. There is still nothing the public can use. |
| Anthropic Claude |
files (.svg, .png, .jpg) | Live | Signed provenance metadata, C2PA standard[1][10][11][13] Supported file types generated by Claude, worldwide. Anthropic's help centre lists 16 models for file credentials: Fable 5.1 and 5, Mythos 5.1 and 5, Opus 5.5, 5, 4.8, 4.7, 4.6 and 4.5, Sonnet 5.5, 5, 4.6 and 4.5, Haiku 5.5 and 4.5. Its developer docs name 14 signed formats, image, video and audio, and say text files, PDFs and office documents are not signed. |
Public C2PA manifests are readable by any C2PA-compatible verifier; Anthropic's developer docs name the open-source c2patool. Anthropic also runs a free one of its own, the Claude Content Checker, which reads the credential in the browser without uploading the file; the help centre began linking it in its 1 Sep 2026 edit. It checks files, not text. Metadata can be stripped by re-encoding. |
| Google Gemini |
text | Live | SynthID Text — token probability modulation[3][4] Gemini app and web experience This row said until 13 Aug 2026 that you could also upload text to Gemini and ask. DeepMind's page does not say that: both detection routes it describes take an image, video or audio file, and neither names text. Whether the Gemini route works on text is unknown to us. |
Open library The scheme is open source with a reference detector on PyPI. Google's hosted SynthID Detector, which DeepMind's page described until Oct 2026 as in testing with journalists and media professionals, is described since an edit read on 8 Oct 2026 as open to everyone, but for image, video and audio files; the page does not say it takes text. |
| Google Imagen, Veo |
image, video | Live | SynthID — imperceptible pixel-level watermark[3] Google generative AI consumer products |
Public Since an edit to DeepMind's SynthID page read on 8 Oct 2026, the SynthID Detector is described as open to everyone for image, video and audio files, checking content made with AI from Google or its partners, named as OpenAI, NVIDIA, Kakao and, soon, Apple. Gemini, Search and Chrome also check an uploaded clip on request. Until that edit the page described the portal as in testing with journalists and media professionals behind a waitlist, and this row read Gated. We have not used the portal ourselves. |
| Google Lyria, NotebookLM |
audio | Live | SynthID audio watermark[3] Lyria music generation, NotebookLM podcast feature |
Public Same SynthID Detector as image and video, described as open to everyone since the Oct 2026 page edit; read Gated until then. |
| OpenAI ChatGPT, ImageGen, Sora, Codex |
image | Live | C2PA Content Credentials plus SynthID watermarking[9][3] Images generated through ChatGPT, Codex or the OpenAI API. Content Credentials since 2024; SynthID added 2026. |
Public OpenAI runs a public verification tool in preview, checking both Content Credentials and SynthID, plus verification API access since 31 July 2026. Limited to OpenAI-generated content. If nothing is detected it deliberately draws no conclusion. Since Oct 2026 DeepMind's SynthID Detector page also names OpenAI among the partners whose AI-made image, video and audio files it checks. |
| OpenAI ChatGPT, OpenAI API |
audio | Live | SynthID watermarking[9][3] Supported audio generated with OpenAI tools, from 31 July 2026 |
Public Covered by the same public verification tool and verification API. DeepMind's SynthID Detector page also names OpenAI among the partners whose files it checks. |
| OpenAI ChatGPT, Codex, OpenAI API |
text | Partial | textGrain, OpenAI's own statistical watermark in the model's word choices (not SynthID), announced 5 Oct 2026[14][15][9] Since 5 Oct 2026 API customers worldwide can opt in for select models; off by default. OpenAI says it will add the watermark to eligible ChatGPT and Codex text in the EU only over the coming weeks, across all plans, and is not making it a global default at launch. Its help centre lists text as 'textGrain watermarks (EU only)'. Which models are 'select' is shown only inside the API settings, which we have not seen. This row read 'None' until 7 Oct 2026, which was correct: OpenAI's provenance page did not mention text until its 5 Oct 2026 update. OpenAI says textGrain is its own scheme, not SynthID, and that it plans to open-source it, without a date. The 5 Oct post says the EU rollout to ChatGPT and Codex is still weeks away, while the help centre table describes EU ChatGPT text as watermarked in the present tense; we follow the post and will move this row when OpenAI states the rollout is complete. |
Gated Applications opened on 5 Oct 2026 for approved researchers and expert organisations, case by case. OpenAI says it is not making the detector public at launch because of missed watermarks and false positives, and that enabling watermarking as an API customer does not grant detector access. Since an edit of 9 Oct 2026 its help centre says text detector checks are free, that qualifying uses are academic and research organisations studying provenance or detection, and that general business applications and academic integrity checks, such as checking student work for AI use, are outside the programme's scope. It publishes figures: at a 1% false positive rate about 80% of 200-token and 95% of 400-token passages were detected; replacing 25% of words with synonyms cut detection to 17%. |
| Meta Meta AI, Llama |
text, image | Unverified | Unknown[5] Unknown Named in reporting as having committed to compliance. We have not located a primary source describing what Meta actually ships. This row is a known gap, not a finding. |
None Unknown. |
| Microsoft Copilot |
text, image | Unverified | Unknown[5] Unknown Named in reporting as having committed to compliance. Primary source not yet located. |
None Unknown. |
How to read this
Status is what the provider ships. Live means the provider itself confirms it. Partial means some models, surfaces or regions only. Unverified means it was reported by third parties and we have not located a primary source. Those rows are gaps in our knowledge, shown rather than hidden, and they are the rows most likely to change.
Can you verify it is the column that matters most and gets the least coverage. A mark that only its author can read gives you nothing as a reader.
The three big providers have answered that question differently, and the spread is the most useful thing on this page. OpenAI ships a public verification tool for its images and audio, and a verification API[9]; for text, it announced its own scheme, textGrain, on 5 October 2026 and gated the detector to approved researchers and expert organisations by application[14]; it says the checks are free and that checking student work for AI use is outside the programme's scope[15]. Google open-sourced its text watermarking scheme with a working detector[4], and since an edit to DeepMind's page read on 8 October 2026 describes its hosted SynthID Detector as open to everyone for image, video and audio files, including partner content from OpenAI, NVIDIA and Kakao[3]. Anthropic shipped the mark, and on 1 September 2026 opened a detection API in private preview to organisations that EU law obliges to check, plus enterprises under the same obligation[10][1]. Marking content and deciding who may check it are separate decisions.
Note the asymmetry by modality, too. Images and audio are comparatively well served, and Anthropic added a free file checker for C2PA credentials in September[1][11]. Text — the modality most people are actually anxious about, because it is what gets students accused — is the one where no member of the public can check anything from anyone except Google[1][14][4].
What changes this table
- Anthropic opening the watermark detection API beyond its private preview. That happened in part on 1 September 2026: the row moved from None to Gated when the API reached organisations obliged to check under EU law[10][1]. It becomes Public only if anyone can run it, and Anthropic says access will widen over time without saying how far[12].
- Primary sources appearing for Meta and Microsoft. OpenAI's appeared on 5 October 2026, when it announced textGrain, and that row moved from None to Partial with a Gated detector[14]. It becomes Live when OpenAI states the EU rollout to ChatGPT and Codex is complete.
- Google opening the SynthID Detector portal beyond journalists and media professionals. That happened by 8 October 2026: DeepMind's page now says access is expanding to everyone for image, video and audio files, and the two Google file rows moved from Gated to Public[3]. Text is still not named as something the portal takes.
Sources
- How Claude marks AI-generated content primary
- SynthID primary
- google-deepmind/synthid-text primary
- Anthropic says it will watermark text generated by its AI models
- Advancing content provenance for a safer, more transparent AI ecosystem primary
- How Claude's text watermark works primary
- Check if files were made with Claude primary
- Introducing Claude Fable 5.1 and Claude Mythos 5.1 primary
- Code execution tool — Content Credentials on generated files primary
- Our approach to EU text provenance rules primary
- Provenance signals in OpenAI-generated content primary
Last verified against primary sources: