Paste a paragraph from Claude into an email and nothing looks different. No odd spacing, no invisible characters, nothing a find-and-replace would catch. But if it came from a recent model such as Opus 5.5 or Sonnet 5.5, the wording itself carries a pattern that someone with the right key can pick up. From November 7, text from the Grok API will carry a watermark of its own.
The timing is not a coincidence. Article 50 of the EU AI Act requires providers of generative AI to mark their output in a machine-readable way, and systems already on the market have until December 2 to comply.
The mark is in the word choice
Neither company adds anything you could strip out. SpaceXSI's watermarking documentation says the watermark works by slightly shaping which tokens the model picks, and adds no characters, markup or hidden text. Anthropic's explainer, published August 14 and updated September 1, says the same of Claude: the text says what it would have said anyway, there are no hidden characters, and there are no extra tokens, so it does not cost more. The technique is a version of SynthID-Text, which Google DeepMind published in *Nature* in 2024.
Anthropic also says the watermark carries nothing that identifies a user, an organization or a chat. It is switched on worldwide, not just in the EU, because Anthropic has no reliable way yet to limit it by region.
Light edits won't shake it off
So far only Anthropic has said how the mark holds up. It describes a pattern that is undetectable to the reader but detectable to anyone who has the key that encodes it. Light editing probably will not remove it completely; a complete rewrite where every word is replaced will. Very short text, heavy edits, and translation or paraphrasing by another tool can also leave nothing to detect. A translation Claude produces itself is still watermarked.
Even a hit says less than you might think. It shows Claude was involved. It does not prove who wrote the text, and it cannot tell "Claude wrote this" apart from "Claude heavily edited this". If you are a teacher hoping to catch an essay, that is a real limit.
Anthropic's help article How Claude marks AI-generated content lists which models carry it on Anthropic's own API and on AWS, Google Cloud and Microsoft Foundry:
- Text watermark now: Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5.
- C2PA Content Credentials on files only, no text watermark yet: Mythos 5, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5.
Models released in the EU on or after August 2 get the watermark at launch. For older ones, Anthropic says it is adding watermarks "over the coming months". Marking covers the API, the Claude apps, Claude Code, Cowork and Claude Tag.
Grok API: on by default from November 7, and opting out is on you
SpaceXSI is the old xAI. It took the name SpaceXAI sometime between June 26 and August 6, and the current terms, dated October 27, already use the present one. An earlier version of those terms and a training-content summary for Grok Imagine Image 2.0 fill in the rest.
Now, the watermark itself. If you run an agent that calls the Grok API every morning to draft a report, its output will be watermarked by default from November 7. That covers grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the grok-4.20 reasoning, non-reasoning and multi-agent models, on both the Responses and Chat Completions APIs, including structured outputs and tool-call arguments, on api.x.ai and every regional endpoint. SpaceXSI says quality and speed are unaffected and you do not need to change your code. To see whether a given response is watermarked, look for the x-watermark: true header.
A team admin can turn it off per model in the SpaceXSI Console, but only after confirming that a legal exception applies and that the team is solely responsible for any marking, disclosure or other legal requirement. Some models may be locked on for your team, and those cannot be turned off from the Console. Every change goes into the audit log and takes effect within 5 minutes.
Detection, SpaceXSI says, is statistical. Longer text gives a stronger signal, and a check comes back watermarked, not watermarked or uncertain. Access to the detection API is by application.
December 2 is the deadline, and not everything needs a mark
The European Commission's FAQ on Article 50 says Article 50 has applied since August 2, 2026. The only grace period is for systems placed on the market before that date, and only for the marking and detectability rule in Article 50(2). Those systems have until December 2, 2026. Fines can reach €15 million or 3% of worldwide annual turnover.
Some output is left out. The Commission's Guidelines on Transparency of AI-Generated Content, which the FAQ cites, list a short string of numbers, symbols or letters; source code; output that only ever travels between machines and gets processed automatically, with no person looking at it; and output used only inside closed-loop industrial and product development work, such as film production, though the final product is not covered by that exception. If the AI is just helping with standard editing, no mark is needed either. The FAQ also says output used between businesses or in industrial settings might get a further, very narrow exemption, if it meets the conditions in the guidelines.
Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 and puts the total at around 190 signatories, itself included. The Commission's Code of Practice page likewise says about 190 companies and organisations had signed by the end of July 2026. Anthropic says other major model developers signed too and will add their own watermarks; this piece has not checked each of them.
You can't check text yourself yet
There is no public text-watermark checker for ordinary users. Anthropic's detection API is in private preview, open to regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and others, and businesses with their own compliance obligations can register interest. SpaceXSI's detection API also requires an application. Third-party detectors such as Pangram don't have Anthropic's key; Anthropic says they analyse writing style, which is a different thing from reading a watermark.
Images are another story. Since October 7, Google's SynthID Detector has been open to everyone worldwide, in English only for now. It checks AI content from Google and partners including OpenAI, NVIDIA, Kakao and Apple, and Google's announcement does not mention text. That is the kind of mark Apple added in iOS 27.2, released October 26, for Photos and Image Playground; the iPhone User Guide says most images created or edited with Apple Intelligence tools carry a hidden SynthID watermark you can check with the SynthID Detector. For files from Claude, the free Claude Content Checker reads C2PA Content Credentials only. It does not look for the text watermark.
Several things are still open: whether text in the Grok app, as opposed to the API, will be watermarked; whether SpaceXSI's watermark uses a key and what copying or editing does to it, which its documentation does not say; when Anthropic's older models will be covered, and whether that happens before December 2; and whether Apple will watermark AI text at all, on which we found no official statement. So don't assume an older model means no watermark. It may get one within months.
For most people there is nothing to do. Claude charges nothing extra for the watermark, and the Grok API documentation mentions no price change. What is worth keeping in mind is the limit: a watermark can link text to a model, not to a person. If you build on the Grok API, put one check on your list for November 7: look for x-watermark: true in your responses, and if your use really falls under a legal exception, see whether the model can be switched off in the Console.