SpaceXSI will start hiding an invisible watermark in text from seven Grok models on November 7, and it will be on by default for every API customer, according to a developer documentation page the company last updated on October 31. Anthropic is extending its own AI text watermark to four older Claude models, according to an email to API customers described by The Decoder. InferenceWire reports the start date as November 3.
Both moves trace back to one rule. Article 50(2) of the EU AI Act requires AI providers to mark generated text in a machine-readable way, and systems already on the market before August 2, 2026 have until December 2 to comply, according to the European Commission’s AI Act Service Desk. The result is that the big chatbot makers now mark text in very different places. Claude does it worldwide, ChatGPT only in the EU, and Grok’s API does it for everyone unless an admin opts out.
- Grok API: watermarked by default from November 7, 2026 on grok-4.7, 4.6, 4.5, 4.3 and three grok-4.20 models. Team admins can opt out per model if they confirm a legal exception applies.
- Claude: models launched since August 2, 2026 are marked at launch. Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 join next, per Anthropic’s email as reported. The Decoder reports customers can’t switch it off.
- ChatGPT: text is watermarked only for users in the EU. On the OpenAI API, watermarking is opt-in and off by default.
- The catch: none of the three text detectors is public, and heavy edits or translation can erase the signal. Very short answers may not carry one at all.
Grok’s watermark is on unless an admin turns it off
SpaceXSI says it watermarks Grok output “to comply with the EU AI Act,” and it applies the mark on every endpoint, including the global api.x.ai address and regional ones, not just for European traffic. The mark covers ordinary replies, structured outputs and tool-call arguments from both the Responses and Chat Completions APIs. Developers don’t send any new parameter. Each response that carries the mark returns an x-watermark: true header.
Opting out is a team-wide setting in the SpaceXSI Console. Admins clear a checkbox for each model, and the change reaches new requests within five minutes. To do it, they have to confirm that “an exception under applicable law applies” to their use and accept responsibility for any marking themselves. Every switch lands in the team audit log, and SpaceXSI says some models may be locked on. It also says the watermark doesn’t change response quality or speed. That is a company claim, and the page offers no test data for it.
Claude already marks its newest models everywhere
Anthropic’s help center page says it has signed the EU’s Code of Practice on marking AI-generated content. Claude models launched on or after August 2, 2026 carry text watermarks from day one, across the API, the Claude apps, Claude Code, Cowork and Claude Tag, “wherever Claude is offered, worldwide.” Files Claude creates also get C2PA Content Credentials. The mark is applied at the model level, so it also shows up when Claude runs on AWS, Google Cloud or Microsoft Foundry.
The email described by The Decoder adds Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, with the Amazon Bedrock rollout possibly taking a few extra days. There’s a catch on two of them. On Opus 4.6 and Sonnet 4.6, the watermark applies only when temperature, top_p and top_k stay at their defaults. “If you rely on the watermark for your own compliance, keep these settings at their defaults,” Anthropic wrote, according to The Decoder. The email also says the mark “contains no information about the user, their organization, or their conversations with Claude.”
ChatGPT watermarks text only in the EU
OpenAI’s help page on provenance signals says ChatGPT text carries its textGrain watermark “in the EU” only. API customers anywhere can turn it on under Organization settings > Data controls > Text provenance, or per project, and pick which models get it. OpenAI says the EU Code of Practice doesn’t require watermarks on outputs shorter than 200 tokens, about 150 words, or on code snippets.
OpenAI’s October 5 announcement is unusually frank about the limits. In its own tests on answers to psychology questions, the detector caught about 80% of 200-token passages and about 95% of 400-token passages at a 1% false-positive rate. On 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. Results by language ranged from 69.0% for Spanish to 42.2% for Romanian.
How an AI text watermark works
None of these marks are hidden characters. OpenAI, Anthropic and SpaceXSI all describe a pattern in word choice. While the model writes, the system nudges which of several plausible next words it picks, and OpenAI says a secret key decides the nudge. A detector holding the key then checks whether the text follows that pattern more often than chance would allow. SpaceXSI’s detector answers watermarked, not watermarked or uncertain, and all three companies say longer text gives a stronger signal.
Google DeepMind built the best-known version, SynthID. Google’s SynthID page says it marks text generated by the Gemini app and web experience, and SynthID Text is available open source for developers. The Decoder reports that Anthropic uses a variant of the SynthID text method. OpenAI says it built textGrain instead to get more control over the trade-off between detectability and variety, and plans to open-source it. Google’s public SynthID Detector checks image, video and audio files.
Who can actually check for the mark
Not the general public, for now. Anthropic’s watermark detection is in private preview for eligible groups such as regulators, media, fact-checkers, researchers and educational organizations. SpaceXSI grants detection access by application. OpenAI limits its text detector to approved research groups and says checking student essays for AI use is outside the program’s scope.
OpenAI and Anthropic both warn that a missing mark proves little. Text written before marking started carries no mark, and heavy paraphrasing or translation can wipe it out. Very short passages often don’t hold enough signal either. Anthropic adds that a found mark means only that Claude may have touched the text, since people use it to proofread and translate their own writing.
What to watch next
The next dates are November 3 for the older Claude models and November 7 for Grok. Anthropic’s help page still lists Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 without text watermarks, so the table there should change first. After that, the December 2 deadline will show whether OpenAI keeps ChatGPT’s text mark EU-only and whether Google says anything about text from the Gemini API.
Frequently asked questions
Does ChatGPT watermark its text?
Only for users in the EU. OpenAI says ChatGPT text there carries its invisible textGrain watermark to comply with the EU AI Act. Outside the EU, ChatGPT text isn’t watermarked, and on the OpenAI API, watermarking is something customers have to switch on.
Can I turn off the Claude or Grok watermark?
Not on Claude. Anthropic applies the mark at the model level, and The Decoder reports customers can’t disable it. On Grok’s API, team admins can opt out per model in the SpaceXSI Console, but they must confirm a legal exception applies, and some models may be locked on.
Can teachers detect an AI text watermark?
Not with Anthropic’s, OpenAI’s or SpaceXSI’s tools right now. None of their text detectors is public. Anthropic’s private preview does include educational organizations, but OpenAI says checking student essays for AI use is outside its detector program, and short or heavily edited text may carry no detectable mark.
Does the watermark change what the AI writes?
It changes which of several similar words the model picks, not the meaning, according to OpenAI, Anthropic and SpaceXSI. OpenAI published benchmark results showing differences within normal run-to-run noise. SpaceXSI and Anthropic state the same but haven’t published comparable test data on their pages.
Sources
- SpaceXSI, “Output Watermarking” (developer docs, last updated October 31, 2026): docs.x.ai
- Anthropic, “How Claude marks AI-generated content” (Claude Help Center): support.claude.com
- The Decoder, “Anthropic’s watermark rollout leaves a loophole on two older Claude models” (Matthias Bastian, October 31, 2026): the-decoder.com
- InferenceWire, “Claude’s Watermark Lives in Word Choice, So Settings Erase It” (October 31, 2026): inferencewire.com
- OpenAI, “Provenance signals in OpenAI-generated content” (OpenAI Help Center): help.openai.com
- OpenAI, EU text provenance announcement (October 5, 2026): openai.com
- Google DeepMind, SynthID: deepmind.google; Google AI for Developers, SynthID Text: ai.google.dev
- European Commission, Code of Practice on marking and labelling of AI-generated content: digital-strategy.ec.europa.eu
- European Commission AI Act Service Desk, implementation timeline: ai-act-service-desk.ec.europa.eu
- Google DeepMind, “From deepfakes to DNA: the science of watermarking AI” (official YouTube channel, October 2026): youtube.com
All sources accessed November 1, 2026.
Last updated: November 1, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
