You've probably used Claude for something in the last month. A draft. A translation. An email you were stuck on. If you did: that text has an invisible signature on it now.

Anthropic announced on August 11 that all new Claude models launched from August 2, 2026 will embed machine-readable watermarks in text and provenance metadata in generated files. Not just for EU users. Everywhere. TechCrunch covered the announcement; Euronews flagged the global-not-just-EU angle.

The move responds to Article 50 of the EU AI Act, which took effect the same day and requires AI providers to mark generated content so that downstream systems can identify it. Anthropic decided not to limit the watermark geographically. Which is either principled or just the path of least engineering complexity — maybe both.

Here's how it actually works.

The text side. Anthropic adapted Google DeepMind's SynthID-Text (think of it as a statistical signature hidden inside word choices — when the model picks between words that work equally well, it subtly favors a pattern that's detectable later with the right key). The signal is imperceptible to a person reading normally. Anthropic is developing a detection API with public keys so any third party — a school, an employer, a platform — can run the check.

One thing that surprised me in the technical details: the watermark travels when you copy and paste. You paste Claude's paragraph into a Google Doc, the mark is still there. That's the design. Whether that's reassuring or unnerving depends on who you are and what you're doing with AI text.

The file side. For images and other supported files (.svg, .png, .jpg), Anthropic attaches C2PA provenance metadata — a cryptographic signature recording what created the file, based on an open standard already supported by major platforms. This is more verifiable than the text watermark and more resistant to stripping. The signature breaks when the file is tampered with.

The watermark does not prove that Claude generated the content, only that it carries Claude's mark. Users often ask Claude to edit, translate, or summarize human-authored text, and the resulting output may carry a mark.

That limitation matters. I'll come back to it in the take.

Source spread

• TechCrunch — Anthropic announces watermarking, Aug 11 [builder] — solid technical coverage with the SynthID connection
• TechCrunch — more details, Aug 15 [builder] — follow-up with detection API specifics and the December deadline
• Euronews — compliance framing, Aug 11 [hype] — emphasizes the global-not-EU-only decision
• The Register — skeptical take, Aug 11 [skeptic] — "sop to EU" framing; focuses on detection gaps and editing edge cases
• BleepingComputer — technical breakdown [builder] — clearest explanation of the SynthID-Text mechanism

Pros & cons

What's real:

• SynthID-Text deployed at scale is genuinely new. This is a Google DeepMind research technique; Anthropic is the first major lab to ship it in production on a consumer product.
• The decision to apply this globally rather than geo-gating it to EU users eliminates the arbitrage that was obvious the moment Article 50 passed. A geographic gate would have just created demand for "the unwatermarked version."
• C2PA for files is mature, open-standard infrastructure. That part actually works and is already compatible with verification tools at major platforms.
• The detection API existing at all is meaningful — it puts the check in the hands of third parties, not just Anthropic.

What deserves a side-eye:

• The text watermark does not prove authorship. If you use Claude to translate a letter you wrote, the resulting text carries a watermark. The mark says "Claude touched this," not "Claude wrote this." Those are very different claims, and the confusion is going to happen at scale.
• Heavy editing strips the marks. The SynthID-Text signal is statistical — it works over enough words in the original pattern. Aggressive rewriting defeats it. The people being caught will be the oblivious ones, not the careful ones.
• Older Claude models have until December 2, 2026 to comply. If you're trying to detect AI content from the last year's worth of Claude usage, there's no watermark to find.

What to do about it

For everyone:

• If you use Claude to write things, the text carries a mark. Understand that heavy editing strips it; light editing doesn't. This affects how you should think about "AI-assisted" versus "AI-generated."
• If you're submitting AI-assisted work somewhere that prohibits it: understand that detection is now technically possible via the watermark. Detection tools still need to integrate Anthropic's API, which takes time — but the window where "use Claude, paste it in, undetectable" was reliably true is narrowing.
• If you receive AI-generated content and care about knowing: the detection API is on its way. It will tell you whether text carries Claude's mark, not whether Claude wrote every word.

For builders:

• The detection API is the one to track for content verification, moderation, or academic integrity tooling. When it ships, it's a new signal worth adding.
• C2PA metadata on Claude-generated images is available now. If you're generating images through the API and care about downstream provenance, the support is there.
• December 2 deadline for older models: historical content from before August 2 carries no watermark regardless of which Claude model produced it. Design accordingly if you're building retrospective detection.

Further reading

• Anthropic watermarking FAQ (PDF) — primary source with technical limitations explicitly stated
• TechCrunch — Anthropic shares more details, Aug 15 — best technical follow-up with December deadline
• BleepingComputer — how it works — clearest SynthID-Text explanation
• The Register — skeptical take — useful pushback on the compliance framing
• C2PA open standard — what Anthropic uses for file provenance metadata