How AI Labs Watermark Text

Google, Anthropic and OpenAI each watermark AI-written text, and xAI is about to start, each drawing the line in a different place. xAI's API documentation, published under the name SpaceXSI, says that from Nov 7, 2026, seven Grok models will mark their text by default on every API endpoint. OpenAI said on Oct 5 that its API stays off by default, and that it would mark ChatGPT and Codex text in the EU only.

The EU deadline behind it

The EU AI Act's Article 50 requires providers of generative AI to mark its output in a machine-readable way. The obligations apply from Aug 2, 2026, and AI systems placed on the market before that date have until Dec 2, 2026. About 190 organizations had signed the EU's Code of Practice on AI-generated content by the end of July, and the EU's list of signatories includes Anthropic, Google and OpenAI. xAI is not on it. Anthropic, OpenAI and xAI each say they watermark text to comply with the Act.

Anthropic marks everywhere, OpenAI in the EU

Anthropic said on Aug 14 that it applies watermarking globally "because we don't yet have a durable way to scope it by region." Its help center lists the surfaces: the API, the Claude apps, Claude Code, Claude Cowork, Claude Tag and the AWS, Google Cloud and Microsoft Foundry clouds. Models launched from Aug 2, 2026 are marked at launch: its post launching Claude Fable 5.1 and Mythos 5.1, which it released on Sep 1, says the Code "required us to add a watermark" to "the outputs of models released after August 2, 2026". Anthropic says it is adding marks to older models. On Nov 1 its table showed text marks on Claude Opus 4.8, Fable 5, Sonnet 5 and Opus 5, which launched before Aug 2, and none yet on models such as Claude Opus 4.7 and Sonnet 4.6.

OpenAI took the opposite approach. Its Oct 5 post says it is "not making text watermarking a global default at launch": it would add the mark to ChatGPT and Codex text in the EU "over the coming weeks", and API customers anywhere can turn it on for select models. Its help center lists text marking as "textGrain watermarks (EU only)" and says ChatGPT text in the EU now carries it.

xAI marks its API, with an opt-out

xAI's page names grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 models, and covers plain text, structured outputs and tool-call arguments. Team admins can turn the mark off per model after confirming "that an exception under applicable law applies to your use", and some models may be locked on. Neither this page nor xAI's consumer terms and FAQs mention a text watermark in the Grok apps. Its Grok FAQ says generated images and videos carry a Grok watermark.

Google's API, per its forum

Google's pages name the Gemini app and web for text watermarking, and its SynthID Detector post says SynthID covers text from its Gemini models "across Google". None of its documentation names the Gemini API. On Google's developer forum, a Google staff member wrote on Aug 19 that "text generated via the Gemini API IS actually SynthID-watermarked", and that the same applies to Google AI Studio and Antigravity. The reply corrected an Aug 5 reply that said API text was not watermarked. Later questions in the thread, on Vertex AI, on when API marking began and on how others can check API text, have no reply from Google.

Who can check

Each lab limits who can check text for its mark. xAI grants detection by application, and OpenAI to approved researchers and expert organizations. Anthropic opens its detection API to eligible organizations such as regulators, media and researchers, and to enterprises that must verify watermarks for their own compliance. Google opened its SynthID Detector, which takes text as well as images, audio and video, to early testers and a waitlist on May 20, 2025. Its SynthID page now says anyone can check an image, video or audio file with it, and names no wider access for text.