Which AI chatbots watermark their text in 2026?

The dates differ by provider. Google has marked Gemini text since 2024. Anthropic marks Claude models launched from 2 August 2026 and is adding marking to earlier ones, and OpenAI began rolling out textGrain in the EU on 5 October 2026. Microsoft and SpaceXAI have published nothing for Copilot or Grok text. The marks are patterns in word choice, with details that differ: Anthropic states that nothing is added to its text and there are no hidden characters, and OpenAI’s technical report describes a keyed signal in token sampling.

Who added a text watermark, and when?

Google’s record is the oldest. TechCrunch reported on 23 October 2024, when Google open-sourced the tool, that SynthID Text had been built into Gemini models since spring 2024, and our Gemini entry dates Google’s announcement to May 2024. Google DeepMind’s SynthID page says the company expanded SynthID to text from the Gemini app and web experience, and it names no other Gemini product. Anthropic’s page of 14 August 2026 describes its own mark as a version of that SynthID-Text approach, while its Help Center page, which describes how Claude marks content, does not use the name.

Anthropic’s Help Center says Claude models launched on or after 2 August 2026 support marking at launch, that marking applies worldwide, and that it is working to add marking to earlier models. A note there says the rollout on Amazon Bedrock will complete by 23 October.

How does the EU timetable fit in?

Article 50 of the AI Act applies from 2 August 2026, according to the European Commission, and the Commission’s page on the code of practice says adherence to the code is voluntary while the Article 50 duties are legal obligations. Anthropic ties its marking to the code: its Help Center says that, consistent with its commitments under the Code, it is adding watermarks to models released before 2 August 2026. Morrison & Foerster counted 82 providers and 152 deployers among the signatories on 9 September 2026.

OpenAI’s rollout follows the same geography. TechCrunch reported that ChatGPT and Codex text in the EU is to carry textGrain over the coming weeks, and that developers anywhere can switch it on through the API for select models. It is off by default, and OpenAI is not making it a global default at launch.

What changes on 2 December 2026?

The Commission’s FAQ, last updated on 29 July 2026, describes a targeted grandfathering rule for Article 50(2) that the EU co-legislators agreed and that would give existing systems until 2 December 2026, still framed as “if adopted”. According to Morrison & Foerster, which wrote on 11 September 2026, providers that already sold AI systems in the EU before 2 August 2026 “have until December 2, 2026” to mark AI-generated and modified content. Each provider decides which of its products that covers, and none of the provider pages read for this article says so for its chatbot.

Formal adoption is not confirmed on the Commission pages read, so the date carries that reservation. The FAQ adds that outputs made available before 2 August 2026 do not need to be marked retroactively.

Three items on provider pages are still in progress. Anthropic says it is adding marking to older Claude models over the coming months. OpenAI’s EU rollout is described as running over the coming weeks from 5 October. Microsoft lists planned text watermarking for select Azure OpenAI models in the gpt-5.1 through gpt-5.6 families. A reader who needs the state of a given model on a given day should check the provider’s page, which is why each entry in our watermark database carries its check date.

Do Copilot and Grok have a text watermark?

Microsoft has published no text watermark for Copilot as of 25 October 2026. Its Learn page on content provenance, dated 17 October 2026, covers Microsoft Foundry and not Copilot. It 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, with no confirmed plans for earlier ones, and its detection website accepts image, audio and video files. Microsoft has not said Copilot text is unmarked either.

SpaceXAI’s Grok FAQ, checked on 25 October 2026, describes a visible watermark on generated images and videos that no setting removes, and mentions nothing for text. It carries no statement that Grok text is unmarked, and an absence of documentation is not a denial in either direction. Both rows can change the day a provider publishes a page.

What can a text watermark tell you?

A text watermark proves little in either direction. Anthropic’s Help Center says a detected mark is “not fully conclusive”, and TechCrunch quotes OpenAI that a missing watermark “does not prove human authorship”. Marks weaken as text changes: TechCrunch reported that replacing 10% of words with synonyms cut detection from about 92% to 66% in one test. Anthropic says light editing probably will not remove its mark completely, while “a complete rewrite where every word is replaced will.” Who can read a mark is the subject of our article “Who can check a text for an AI watermark?”, and how textGrain works is the subject of “What the ChatGPT text watermark is, and what Grok has instead”.

Hidden characters are a separate question. Zero-width spaces and narrow no-break spaces come from editors, web pages and clipboards, and a scan for them says nothing about any mark in this timeline. Our invisible character detector finds those code points, and no scanner finds a statistical watermark. Our textGrain article explains why the narrow space in ChatGPT answers was never the mark.