Anthropic, OpenAI and xAI Are All Watermarking AI Text Now — Here’s How It Breaks

Last updated: 2 November 2026. Figures and technical claims below are sourced to the primary publications linked inline — company documentation, official help pages and the EU’s own Article 50 guidance — with any self-reported company figures flagged as such.

The 60-second version

• xAI’s own documentation confirms that, from 7 November 2026, every text response from seven Grok models will carry an invisible statistical watermark by default, with opt-out limited to admins who confirm a “legal exception” applies.
• Anthropic’s help centre confirms nine Claude models already watermark text, with Content Credentials (C2PA) attached to generated files; reporting from The Decoder, citing an Anthropic customer email, says four more models follow from 3 November — though two of them, Opus 4.6 and Sonnet 4.6, only watermark output when developers leave sampling settings at their defaults.
• OpenAI got there first, in October, with a system it calls textGrain — but OpenAI’s own published testing shows that swapping one word in four for a synonym cuts detection from roughly 92% to about 17%.
• All three moves trace back to the same regulatory trigger: Article 50(2) of the EU AI Act and the voluntary Code of Practice on Transparency of AI-Generated Content that each lab has signed.
• None of the three systems gives an ordinary reader, journalist or platform a way to check a piece of text themselves — detection tools are gated to “approved” researchers, regulators and enterprises.

Key dates and numbers

What xAI, Anthropic and OpenAI have actually shipped

Three of the largest model providers have now built, and in two cases switched on, invisible watermarking for AI-generated text — and they did it within about four weeks of each other. xAI — a SpaceX subsidiary since February 2026, now in the middle of rebranding toward “SpaceX Super Intelligence,” though its developer docs still sit at docs.x.ai — describes, in its own documentation, a system that “slightly shapes which tokens the model picks, so it adds no characters, markup, or hidden text.” From 7 November, it applies by default to every text-generation call on the Grok API — the Responses API, the Chat Completions API, structured outputs and even tool-call arguments — across seven named models. A response header, x-watermark: true, tells a caller whether a given output was marked; anything else, including a missing header, means it wasn’t. Turning the mark off requires a team administrator to go into the xAI Console, tick a box for each model, and confirm — in writing, logged in the audit trail — that a legal exception applies to their use.

Anthropic moved first among the three, and has been doing this in stages since August. Its official help-centre page states that Anthropic has signed the EU’s Code of Practice on Transparency of AI-Generated Content — the voluntary instrument tied to Article 50(2) — and that Claude models launched from 2 August 2026 onward ship with watermarking at launch. As of this writing, nine Claude models — including Opus 5.5 and Sonnet 5.5 — carry the mark; seven older ones, among them Mythos 5, Opus 4.6 and Sonnet 4.6, do not yet appear on the “yes” side of Anthropic’s own table. That is where the reporting gets ahead of the documentation: The Decoder reports that Anthropic emailed API customers saying four of those older models would gain the watermark from 3 November — a date the public page had not yet caught up with at the time of writing.

OpenAI was first to publish, if not first to switch anything on everywhere. Its system, textGrain, went live as a default for ChatGPT and Codex users inside the EU “over the coming weeks,” while remaining an opt-in, off-by-default feature for API customers globally — a split that Brussels’ own enforcement posture toward OpenAI and Anthropic makes easy to understand: the obligation is European, so the product behaviour is European, and everyone outside the bloc gets the feature only if they ask for it.

Why a Code of Practice, not a court order, is doing the work

Nothing here was forced by a ruling or a fine. Article 50(2) of the AI Act requires providers to make synthetic audio, image, video and text “marked in a machine-readable format” and “detectable as artificially generated,” using methods that are “effective, interoperable, robust and reliable” — but it leaves the specific technique to providers. The Commission’s July 2026 guidelines and the voluntary Code of Practice on Transparency of AI-Generated Content fill that gap, and — per the Orrick law firm’s summary of the Code — specifically call for “imperceptible watermark embedded within the content” as a mandatory sub-measure, alongside signed metadata, with interoperable detection due by February 2027. Signing the Code is voluntary, and the guidance is explicit that “adherence does not constitute conclusive evidence of compliance” — the labs that signed are betting a documented rollout is the cheapest way to show good faith, not a guarantee against enforcement. Brussels has already shown it will open formal probes into model-makers’ AI Act obligations, as its case against OpenAI and Anthropic over separate provisions shows, so the incentive to visibly comply with Article 50 ahead of any inquiry is real.

The gap between “watermarked” and “detectable”

This is the part most coverage of these rollouts has skipped past. A watermark that can be read only by the company that applied it, and only by a short list of parties it chooses to grant access to, is not the same thing as a mechanism the public, a newsroom or a rival platform can use to verify text. All three systems keep detection gated: Anthropic’s detector is “in private preview” for regulators, law enforcement, media and fact-checking organisations who apply through a form; xAI grants access “by application” through a Google Form; OpenAI’s detector access “will initially be limited to approved researchers and expert organizations.” None has shipped a public checker comparable to what publishers increasingly need for moderation and sourcing — a gap Apple Music ran into from the other direction, when it tried to mandate AI disclosure and found the label only as honest as the upload process behind it.

The second gap is robustness, and here OpenAI’s own published numbers are the most useful data point in this story precisely because the company released them itself rather than having a third party dig them up. On a 400-token passage, OpenAI’s testing found detection confidence at roughly 92% with no edits, dropping to about 66% once 10% of the words were swapped for synonyms, and down to about 17% once a quarter of the words were swapped — well within what a human editor, or another AI model asked to “rephrase this,” would do routinely. Mathematical and technical content, where word choice is more constrained, showed even weaker detection. Anthropic’s situation is different but not obviously better: on two of the models due to gain watermarking, Opus 4.6 and Sonnet 4.6, The Decoder reports the mark only appears when a developer leaves temperature, top_p and top_k at their defaults — parameters that plenty of production API integrations already override for reasons that have nothing to do with evading detection. xAI’s documentation does not publish comparable robustness figures at all, which makes it the least transparent of the three on exactly the question that matters most.

What this means if you build or publish with these models

For anyone integrating these APIs, or running a platform that has to make calls about what’s AI-generated, three things follow directly. First, don’t treat a watermark header as proof of anything on its own: x-watermark: true tells you a vendor’s detector would currently flag that specific output, not that the text stays traceable once copied, edited or run through another model — which remains the scenario that actually drives misinformation risk. Second, if your compliance programme touches the EU and uses the Claude or Grok APIs, check your team’s watermarking settings now rather than after an audit — xAI’s opt-out requires an affirmative, logged legal justification, and not knowing the default changed on 7 November won’t be a comfortable position to explain later. Third, if you’re building AI-content-detection into your own platform, layer it — C2PA metadata, vendor watermarks where available, ordinary provenance signals like publishing history — because none of these three systems is individually robust enough to carry that weight alone, a gap that echoes the one Anthropic’s own IPO filing acknowledged between safety work done and safety work still needed.

What we still don’t know

• Anthropic’s help-centre page had not been updated, as of this writing, to confirm the 3 November rollout that The Decoder reported from a customer email — treat that date and model list as reported, not yet officially confirmed in public documentation.
• Neither xAI nor Anthropic has published robustness figures comparable to OpenAI’s; we don’t know how their watermarks hold up against paraphrasing, translation or adversarial rewriting.
• Google DeepMind, Meta and Mistral have not published equivalent text-watermarking commitments tied to this deadline, so it’s unclear whether coverage across the industry will converge or stay uneven.
• Whether the Code’s February 2027 interoperable-detection deadline produces one shared system, rather than several separate gated tools, is genuinely open.
• No regulator has said what counts as sufficient compliance in practice, since adherence to the voluntary Code explicitly does not itself prove compliance with the binding Article 50 obligation.

FAQ

Does this affect me if I just use ChatGPT, Claude or Grok casually?
Mostly not yet, unless you’re in the EU. OpenAI’s default rollout is EU-only for now; Anthropic’s and xAI’s watermarking is applied at the model and API level globally, but it changes nothing about what the text looks like or how it reads — the whole point is that it’s imperceptible.

Can someone actually prove a piece of text came from one of these models?
Only if they have access to that company’s detector, and all three currently restrict access to approved regulators, researchers, fact-checkers or enterprise compliance teams. There is no public tool today where a reader can paste in text and get a reliable answer.

Is the watermark easy to remove?
Based on OpenAI’s own published data, modest paraphrasing substantially weakens detection, and Anthropic’s reporting shows two models’ watermarks depend on using default sampling settings. Treat all three as fragile against deliberate evasion, not as a forensic guarantee.

Why is this happening now, specifically?
Because Article 50(2) of the EU AI Act became applicable on 2 August 2026, and each company has signed the voluntary Code of Practice that spells out watermarking as the expected way to comply — so the rollouts are arriving in a tight cluster as labs try to be visibly compliant ahead of further enforcement activity.

Sources

• xAI — Output Watermarking documentation (primary)
• Anthropic — How Claude marks AI-generated content (primary)
• OpenAI — textGrain / EU text provenance announcement (primary)
• The Decoder — “Anthropic’s watermark rollout leaves a loophole on two older Claude models”
• Mixed News — reporting on OpenAI’s textGrain robustness figures
• Orrick — EU AI Act Article 50 transparency obligations, summary

Please follow and like us: