When we say on this page that a company “watermarks text” we do not mean that you can test for it. Every text watermark that has been put into use so far is keyed, so reading one takes the secret key, and the company that generated the text is the party holding it, which leaves the test in their hands and not in yours, or your school’s, or your employer’s, and not in ours either. There is no provider text-detection API that the public can call today, and if you want to know why it was built that way our watermark explainer covers it.

Live adapter matrix

Provider states as reviewed on August 16, 2026 (served live by the detection service):

Provider State What that means today
Google announced SynthID-Text published by Google DeepMind (Nature 2024, open-sourced detector code), and deployed in Gemini — but no public verification API is available for third parties, so no adapter can go live.
Anthropic announced Anthropic announced (Aug 11-14, 2026) that Claude models launched after Aug 2, 2026 watermark their text output; its detection API is announced but not yet public, so no adapter can go live.
OpenAI unavailable OpenAI built a text watermarking method internally (reported 2024) but has not released it or any verification API; pending.
Meta unavailable Meta published watermarking research (e.g. video/audio; Three Bricks for text) but ships no public text watermark verification; pending.
Mistral unavailable No announced text watermarking or verification API from Mistral; pending.
Microsoft unavailable Microsoft backs C2PA for media provenance but exposes no text watermark verification API; pending.
xAI unavailable No announced text watermarking from xAI.

The table has three words in it that come from our own software, so we should say what each of them means. A row that says live means we have a working connection to that company’s detection route and your scan is being run through it. A row that says announced means the company has said that it marks its text, or that it will, but it has not given us or the public a way in yet, so we can do nothing with the mark even though we know it is there. And a row that says unavailable means we could not find a mark or a route at all, which for most of the companies on the list is where things stand as we write this. The table is read from the same place our scanner reads it from, which is the service’s status feed, and none of it is typed by hand, so if a row is wrong here the scanner has the same wrong row.

What we can do, and what anybody outside those companies can do, is look at the statistics of the writing and give you a probability with a published error rate behind it, and say plainly that it is a probability. If somebody tells you they can read a provider’s mark out of a piece of text, they are either mistaken or selling something.

What to do while the column says pending

What settles an argument today is what settled it before any of this existed, which is to look at how the writing was made. Drafts, version history and notes are the best evidence there is, and they will still be the best evidence after the companies open up their detection routes, since a mark in the text tells you that a model wrote some words and tells you nothing about who typed them into the essay or why. Our own scan gives you a probability with an error rate you can read on the evidence page, and it belongs in that picture alongside the drafts and the history. If you have already been accused on the strength of a score, the page for people in that position goes through what to keep and what to ask for.

Provider status

Google has been at this the longest and there is still no public way to test for its mark. It has put the SynthID-Text mark on the text that comes out of the Gemini app and the web version since 2024, and that October it published the method in Nature, so if you want to you can read how Google’s mark works [Google DeepMind; Dathathri et al., Nature, Oct 2024]. Detection has stayed in Google’s hands. The SynthID Detector portal was announced on May 20, 2025 and is still behind a waitlist for journalists, media professionals and researchers, the text part of it is not live, and the Content Detection API that Google announced for businesses in May 2026 takes images only [Google blog, May 2025; DeepMind SynthID page; InfoQ, May 2026 — accessed Aug 16, 2026].

Anthropic is new enough here that we would come back to it in a month or two. Between August 11 and 14, 2026 the company said that the Claude models it had launched after August 2 of that year were now putting a mark in their text everywhere, and by everywhere it meant the API, Claude Code and the versions you get through AWS, GCP and Azure, using the same SynthID-Text approach that Google uses, with the older models to follow “over the coming months.” Anthropic’s people also said there would be a detection API that would “estimate the likelihood of Claude’s involvement,” and we would like to see it, but so far there are no docs for it, no waitlist and no date, so for now Claude leaves everybody outside the company exactly where the others do [Anthropic, Aug 2026; TechCrunch, Aug 11, 2026].

OpenAI is the one people ask us about most, since it makes ChatGPT, and it has shipped nothing for text. It signed the EU Transparency Code of Practice in July 2026 and said its goal was “to expand provenance signals to all modalities including text,” and nothing has come out yet. OpenAI also built a text watermark of its own somewhere around 2022 to 2024 and then decided not to release it, and the reasons OpenAI’s people gave at the time were false positives and what it would cost them against the competition. OpenAI’s pictures are a different story, since they have carried C2PA metadata for a while and, since May 2026, SynthID as well [CityAM, Aug 12, 2026; InfoQ, May 2026; OpenAI help center — accessed Aug 16, 2026].

Meta and Mistral are both pending, and for the same reason. Each of them signed the Code of Practice in July 2026, and when we went looking on August 16, 2026 we could not find a shipped text watermark from either one, and we looked fairly hard, since we would have liked to find one. Meta’s published watermarking work, which is Stable Signature and the video and audio marks in Meta AI, was aimed at pictures and sound, and none of it puts a mark in text [EC signatory list, Aug 2026].

Microsoft has done a lot on this and none of it is in the text, since its approach has been metadata. It signed the Code, a picture you make with Copilot already gets C2PA Content Credentials on it, which is the kind of mark you can read for yourself with our file tab, one of Copilot’s audio overviews carries an audio watermark, and anything AI made in Microsoft 365 gets a note attached saying which model, which app and when. What it does not have is a mark hidden in the words themselves, which is the thing you would want if you were trying to test an essay, so for text we have it down as pending [Microsoft Learn / Support, accessed Aug 16, 2026].

xAI has committed to nothing and we do not know what it will do. It did not sign the Code of Practice, and from what was reported it was the only big language-model maker that declined to, and it has said nothing about plans to watermark. That does not get it out of the marking obligation in the EU AI Act, which still applies to any output of xAI’s that reaches the EU, signed or not, so at some point it is going to have to do something [EC signatory list; CityAM, Aug 12, 2026].

Why we expect the “pending” column to get shorter

Two dates in the EU’s rules are going to force the question, like it or not. The first is December 2, 2026, which is when systems that were already on the market have to start marking their text output in a way a machine can read, since that is the day the grace period in the AI Omnibus runs out. The second is February 2, 2027, and by then the companies that signed the Code have to give the rest of us some way to test for their marks. That could be a public access point, or a signpost inside the content itself, or something run by a consortium, and any of those would let people outside the company look for the marks themselves, where today we all have to take the company’s word for it [ReedSmith; TechPolicy.Press; Stibbe — accessed Aug 16, 2026].

The day one of these companies opens up a way to detect its mark, we connect to it, and this page changes, and you do not have to do anything for that to happen. You can come back and look whenever you like, or you can just run a scan, since the technical details on every scan result state which provider watermark checks are live, and they are read from the same live status feed as the table further up this page, so the two of them cannot disagree with each other.

Sources

  1. Google DeepMind, SynthID page, deepmind.google/models/synthid/ (accessed Aug 16, 2026); Dathathri et al., Nature (Oct 2024); Google blog, SynthID Detector announcement (May 20, 2025); InfoQ, AI Content Detection API (May 2026)
  2. Anthropic, Claude text watermark announcement (Aug 2026); TechCrunch (Aug 11 and 15, 2026)
  3. European Commission, Code of Practice signatory list (updated Aug 12, 2026)
  4. CityAM (Aug 12, 2026); OpenAI help center art. 8912793 (accessed Aug 16, 2026)
  5. Microsoft Learn, Copilot watermarks documentation (accessed Aug 16, 2026)
  6. ReedSmith; TechPolicy.Press; Stibbe — analyses of the Code's detection and interoperability deadlines (accessed Aug 16, 2026)