Grok API text gets hidden AI watermarks by default from Nov. 7

Text from SpaceXSI’s Grok API will carry an invisible watermark by default starting Nov. 7. The company says it is doing this “to comply with the EU AI Act,” according to its developer docs. Anthropic already marks text from its newest Claude models under the same law, its Help Center says.

The rule in plain English

Article 50 of the EU AI Act requires providers of generative AI to mark output “in a machine-readable format” so it is “detectable as artificially generated or manipulated,” the European Commission says. Those obligations are “applicable from 2 August 2026.” A voluntary code of practice, drawn up by independent experts and facilitated by the EU’s AI Office, helps companies show they comply. “Even though adherence to the code is voluntary, the transparency requirements under article 50 of the AI Act are legal obligations,” the Commission says. About 190 organizations had signed by the end of July, it says.

What SpaceXSI says

The watermark covers text from seven models, including grok-4.7, grok-4.6 and grok-4.5. It applies to the Responses and Chat Completions APIs, on every endpoint. It covers “ordinary text responses, structured outputs, and tool call arguments.” A response header, “x-watermark: true,” shows the mark was applied.

The docs say the mark “slightly shapes which tokens the model picks” and adds no hidden characters. “Detection is statistical,” and “longer text gives a stronger signal.” Detection access is “granted by application.”

Team admins can opt out per model in the Console. Opting out asks customers to confirm “that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement.” Some models “may be locked on.” Every change is logged.

What Anthropic says

Anthropic says it signed the Article 50(2) code. Claude models launched in the EU on or after Aug. 2, 2026, mark output at launch. It is “working to add marking support for earlier Claude models.” Marks apply in the API, Claude, Claude Code, Claude Cowork and Claude Tag, “wherever Claude is offered, worldwide,” and through AWS, Google Cloud and Microsoft Foundry. Generated files get C2PA Content Credentials where supported. The page does not describe an opt-out.

Detection is in private preview for regulators, law enforcement, media, fact-checkers, independent researchers, educational groups, EU civil society groups and similarly obligated enterprises. Anthropic warns a mark “is not fully conclusive.” Heavily edited, translated or very short text may carry no detectable mark.

Our read

This is compliance, not a breakthrough. Two labs are marking ordinary AI text because Brussels requires it. That matters well beyond Europe: Anthropic applies its marks worldwide, so U.S. users get them too.

Does it work? By the companies’ own accounts, only partly. Both say the marks are harder to detect in short text, and Anthropic says heavily edited or translated text may carry no detectable mark. A missing mark doesn’t mean a person wrote it, Anthropic says. That limits the value for schools. A teacher holding a three-sentence answer gets little from a probabilistic score, and neither lab offers public detection of text watermarks. The keys stay with the labs.

For publishers and fact-checkers, the marks add one more signal, if they win detection access. The design also shifts risk. SpaceXSI’s opt-out puts the legal duty on the customer. Founders building on these APIs inherit decisions about disclosure they may not have planned for.

What’s next

SpaceXSI’s default starts Nov. 7. Anthropic says it will publish “more detailed technical guidance” and expand detection access “over time.” Companies that comply by other means will have to show their measures are adequate, which market surveillance authorities will assess case by case, the Commission says.

Sources: SpaceXSI, Output Watermarking; Claude Help Center, How Claude marks AI-generated content; European Commission, Code of Practice on Transparency of AI-generated Content