Who can check a text for an AI watermark?

Text watermarks now exist at several AI providers, and the people who can check a passage for one are few. A text watermark is a statistical pattern in the words a model picks, readable only by a detector that holds the provider’s secret key. Pasting a text into a website cannot read it unless the site holds that key.

Why can’t anyone just paste a text into a checker?

A text watermark carries no visible or hidden character, so there is nothing for a scanner to find. Anthropic describes its mark as a version of the SynthID-Text approach: a secret key and the preceding words steer the choice among near-equivalent words, and detection asks whether a text’s word choices match what the key would produce. OpenAI’s technical report for textGrain says its detector “requires only the generated text and the secret key”. Without the key, the pattern cannot be measured.

Google’s developer documentation shows what that means in practice for SynthID Text. Detection needs a trained detector, a watermarking configuration stored privately, and two thresholds that trade false positives against false negatives. The owner of a detector then picks one of three exposures: fully private, semi-private through an API, or public. Of the providers below, those that document a text watermark keep its detector at the private or semi-private level, and none documents a public text checker.

Which providers watermark text, and who can check it?

Who can run the OpenAI and Anthropic detectors?

OpenAI’s textGrain detector is open to a short list. TechCrunch reported on 5 October 2026 that OpenAI is providing “initial detector access only to approved researchers and expert organizations”. The rollout is phased, and our textGrain article covers it. Our article for teachers covers what the limits mean for a school.

Anthropic’s detection is also gated, with a wider list. Its page, published on 14 August 2026 and updated on 1 September, says the detection API is in private preview for regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises that must verify watermarks for their own EU compliance. Access is expanding, and interested parties can register through a form on the page. The page also warns that a mark can only show that Claude was likely involved at some point.

Does Google’s SynthID Detector read text?

Google’s SynthID Detector is public since 7 October 2026, and its announcement describes a check on “an image, video, or audio file” made with AI from Google or its partners. The post names OpenAI, NVIDIA and Kakao among the partners, with Apple to follow, and says the tool is available globally in English. It does not list text. The Google DeepMind page on SynthID Text, meanwhile, says the company has expanded SynthID to watermarking and identifying text from the Gemini app and web experience, and it describes no way for the public to check a passage.

Google’s first announcement of the portal, on 20 May 2025, said “Video and text watermark detection will be rolled out in the coming weeks.” The later post lists three media types instead. SynthID Text is also open source, with a production implementation in Hugging Face Transformers from version 4.46.0 and a Bayesian detector provided. That lets a developer check text from a model they watermarked themselves, and it gives no one a way to check a Gemini answer.

What about Microsoft Copilot and Grok?

Microsoft documents text watermarks only through the model provider. Its Foundry page, dated 16 October 2026, says text provenance support varies by model and that, where it exists, the underlying model provides the invisible watermarking. It lists planned support for select Azure OpenAI models in the gpt-5.1 through gpt-5.6 families and points to Anthropic’s guidance for Claude models. Its Microsoft 365 page on watermarks covers images, video and audio, and the content provenance detection website accepts image, audio and video files. For Copilot text, no watermark is documented.

For Grok the record is thinner. SpaceXAI is the company behind Grok. The Grok FAQ describes a visible watermark on generated images and videos and says nothing about text or a detection tool. City AM reported on 12 August 2026 that xAI, now SpaceXAI, did not sign the EU code of practice on transparency of AI-generated content. Not signing a voluntary code says nothing about what Grok does to its text, so the table reads “Not documented” and no more.

What can a reader check without a key?

A reader can check the text for hidden characters, and that is a separate question from a watermark. Our invisible character detector looks for eight code points, such as the zero-width space and the byte-order mark, and highlights each one. Those characters come from editors, web pages and clipboards as often as from a model, so a hit points to a text-handling step and a clean result says nothing about textGrain, Claude’s mark or SynthID Text.

Style detectors answer a third question. Our AI detector reports how closely a passage reads like AI-written prose and quotes the phrases behind the reading. No tool of that kind reads a keyed watermark. A site that reports a “Claude watermark” or “Gemini watermark” verdict for pasted text owes readers an account of its access, because none of the documents above describes public access for text. Before trusting a result, ask whether the page was given the key or only your text.

Access is moving: Anthropic says it will expand access over time, and OpenAI describes its detector access as initial. Each entry in our watermark database carries the date its facts were last checked, and this table reflects the sources named below as read on 17 October 2026.