Anthropic and xAI are expanding invisible watermarking for AI-generated text in the same week, and both companies point to the European Union’s AI Act as the main driver. Anthropic said outputs from four older Claude models will carry the watermark starting Nov. 3, while xAI’s developer documentation says seven Grok models will have watermarking turned on by default for API text output from Nov. 7.
Anthropic extends watermarking to older Claude models
According to the report, Anthropic notified users on Oct. 30 that it was widening the scope of its watermarking system. Starting Nov. 3, outputs from Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 will also include the invisible text watermark. Anthropic had already enabled the same system on its latest models in August, and this change extends it to older versions.
Anthropic’s support page says the company has signed the code of practice on transparency for AI-generated content under Article 50(2) of the EU AI Act. Under that commitment, Anthropic said it would add watermarking to outputs from models released before Aug. 2, 2026. The support page itself does not list Nov. 3 as the effective date. The report says that timing currently comes from Mashable’s reporting and customer email references.
EU AI Act Article 50 is central to the rollout
The European Commission’s website says the transparency obligations under Article 50 of the AI Act apply from Aug. 2, 2026. Providers of generative AI systems must ensure that audio, image, video, and text outputs carry machine-readable markings and can be detected as AI-generated. The related code of practice is voluntary to sign, but the transparency requirement in Article 50 is a legal obligation.
The report also notes that Aug. 2 is the date Anthropic uses on its support page to separate older models that need retroactive labeling.
Watermarks are embedded in word choice and may survive editing
Anthropic says the watermark is inserted directly into the content as the model generates text. Users cannot see it, and the company says it does not change the meaning, quality, or readability of the response. Because the watermark is part of the text itself, it remains after copy and paste and may still persist after partial edits.
Mashable cited Anthropic as saying the system works by leaving “subtle patterns in word choice.” Anthropic also says the watermark does not contain any user or organization identity information. Its support page adds that the watermark is applied at the model level, so text from any Claude product will carry it. For files such as images, Anthropic uses C2PA content credentials, which are signature data attached to files to record provenance.
Text detection is limited to approved groups for now
People receiving watermarked text cannot inspect it on their own at this stage. Anthropic’s support page says text watermark detection remains in a private preview and is available only to groups covered by EU legal requirements. That list includes regulators, law enforcement agencies, media organizations, fact-checkers, independent researchers, educational institutions, and EU civil society groups. Companies with similar compliance review obligations can also use it.
Anthropic said it plans to widen access to the detection API over time, and interested parties can currently register through a form. C2PA credentials attached to files can be checked through the free Claude Content Checker.
Pete Sena wrote in a post that the watermark does not reveal who the user is and only shows that the text passed through Claude. He also wrote, “Even a quick proofread counts.” Anthropic’s support page, though, describes the label as applying to content Claude “generated and processed,” without separate detail for uses such as proofreading or translation.
xAI sets default watermarking for seven Grok models
xAI’s developer documentation says that, from Nov. 7, 2026, text produced through the API by seven Grok models will be watermarked by default. The list includes grok-4.7, grok-4.6, grok-4.5, grok-4.3, and three versions of grok-4.20: reasoning, non-reasoning, and multi-agent.
The scope covers both the Responses API and the Chat Completions API. Standard text replies, structured outputs, and tool call parameters are all included. The report also says the company name in the documentation has been changed to SpaceXSI, and the documents state that watermarking is being added to comply with the EU AI Act.
xAI says the system affects token selection slightly
The report says both companies are approaching watermarking through word choice or token selection. xAI’s documentation says the watermark is applied during generation and slightly affects which tokens the model selects. It does not add any extra characters, markers, or hidden text to the response. Developers can inspect the x-watermark response header, and a value of true means the output was watermarked. xAI says the feature does not affect response quality or speed.
xAI’s detection method is statistical and returns one of three outcomes: watermarked, not watermarked, or inconclusive. The signal gets stronger as the text becomes longer. Access to the detection API requires a separate application. Team administrators can disable watermarking model by model in the dashboard, but they must confirm that their use case falls under a legal exception and take responsibility for any labeling and disclosure obligations. Some models are locked with watermarking enabled and cannot be turned off manually.
Both companies say more guidance is coming
Anthropic’s support page says it will update its documentation and publish more detailed technical guidance later. xAI’s watermarking goes live on Nov. 7, at which point API text outputs from the seven listed Grok models will carry the marker by default.
