The date does the talking. Anthropic's support article, published today, says that Claude models launched on or after August 2, 2026 weave an imperceptible watermark directly into the text they generate, and attach digitally signed C2PA provenance metadata to supported file outputs — .svg, .png, and .jpg. August 2, 2026 happens to be the exact day the EU AI Act's Article 50 transparency obligations started to apply — including the duty for providers of generative systems to mark AI-generated content in a machine-readable way. The Register's headline was less diplomatic than the support article: "watermarks to help discern AI slop in sop to EU." Anthropic, for its part, says the marking applies worldwide, not just to European users — which is what compliance usually looks like when maintaining two output pipelines costs more than complying everywhere.
What actually ships
Two separate mechanisms, and the distinction matters:
Text: a statistical watermark. The watermark is woven into the generated text itself — invisible to the reader, present across the API, the Claude apps, and third-party platforms that serve these models. Because it lives in the word choices rather than in metadata, it survives copy-paste, and Anthropic says it expects some marks to survive light editing too. This is the well-studied "green-list" family of schemes: bias token sampling toward a keyed subset, then test statistically for that bias later. The mark is a property of the text, not a tag on it.
Files: signed C2PA metadata. Image and SVG outputs carry a cryptographically signed provenance manifest under the C2PA standard — the same coalition framework OpenAI adopted for DALL·E images back in 2024. The signature, in Anthropic's words, "signals that a file was processed by Claude and lets you detect whether the file has been tampered with." Note the verb: processed by, not generated by — a deliberately narrow claim, and the correct one.
What the marks prove — and what they can't
The honest reading of both mechanisms is that they are presence tests, not absence tests, and weak ones in different ways:
- C2PA metadata is trivially removable. Screenshot the image, re-save it through any tool that strips metadata, and the manifest is gone. C2PA's own design acknowledges this: a valid signature proves provenance; a missing signature proves nothing. It answers "did this verified file come through Claude intact?" — useful for institutions checking documents — not "is this random image AI-generated?"
- The text watermark degrades under rewriting. Anthropic itself says heavy editing or translation may render it undetectable. Paraphrase a paragraph through another model and the statistical signal washes out. And short outputs — a sentence, a headline — carry too few tokens for the test to reach significance in the first place.
- Detection is not attribution. Finding the mark tells you the text may have passed through Claude at some point — quoted, pasted, partially edited — not that Claude authored the document you're holding. The support article is careful on this; secondary coverage mostly isn't.
- It only marks the marked. The watermark applies to models launched on or after the cutoff. Everything generated by earlier Claude models, every other lab's unwatermarked model, and every open-weights model anyone runs locally produces text with no mark at all. A negative result is meaningless in a world where most of the models shipping are open weights.
And one asymmetry worth sitting with: there is no public detector. Anthropic has not shipped a tool that lets a teacher, an editor, or a regulator check a suspect paragraph. The watermark currently satisfies the letter of "machine-readable marking" — the mark exists, machines could read it — while the only machines that can actually read it are Anthropic's. Whether Article 50 and the Commission's Code of Practice on Transparency end up requiring verifiability by others, not just markedness, is the genuinely open legal question here.
The legislative context, precisely
To place this on the timeline from our Digital Omnibus post: the Omnibus deferred the AI Act's high-risk obligations to December 2027, but Article 50 transparency was exactly the part that did not move — it applied on 2 August 2026, with enforcement live at national and EU level. The one softening is transitional: providers of synthetic-content systems already on the market by that date get until 2 December 2026 to comply with the machine-readable marking duty specifically. Read Anthropic's cutoff against that carve-out and the design is legible: models launched after August 2 get no grace period, so they watermark from day one; the older models still serving traffic sit inside the transitional window. The rollout tracks the regulation clause by clause — which is the strongest evidence that the regulation, not product conviction, set the schedule.
That's not a criticism, exactly. It's the transparency provisions doing what they were written to do: converting a voluntary-commitment culture (recall the 2023 White House pledges, which produced OpenAI's C2PA adoption and little else) into a dated legal obligation that shipped on the date.
What to expect next
- Watch for a public detector — or its conspicuous absence. The gap between "we mark it" and "you can check it" is where the compliance question lives. If a detection API appears, watch who gets access: everyone, or vetted institutions.
- Watch the other labs' cutoff dates. If Google, OpenAI, and Mistral announce text watermarking with effective dates clustered around August–December 2026, the EU clock — not a research breakthrough — is what made 2026 the year of the watermark.
- Watch the removal arms race, which started on day one. Watermark-remover tools were being marketed within hours of the announcement. The interesting question isn't whether paraphrase attacks work — they do — but whether detectable-after-light-editing turns out to cover enough real-world misuse (spam farms, essay mills copy-pasting at scale) to matter anyway.
- Watch open weights. No marking scheme survives the model weights being public — a local model samples however its operator wants. If regulators notice that the marking duty only binds the labs easiest to regulate, the pressure lands next on open-weights releases, and that's a much uglier fight.
References: Anthropic — How Claude marks AI-generated content · Interesting Engineering — Copy-paste no more · BigGo — AI-Generated Text Remains Traceable After Copy-Paste · Business Standard — How Anthropic marks AI-generated text · The Register — Anthropic pledges to embed watermarks in sop to EU · The Register — OpenAI's 2024 C2PA adoption · related coverage: The AI Act's Compliance Cliff Just Got Deferred — Except Where It Didn't · Frontier Arcade: trends & predictions