Anthropic to Watermark Claude-Generated Text and Files Under EU AI Act
Anthropic says new Claude models will embed invisible text watermarks and C2PA metadata to comply with the EU's AI transparency rules.
Conversation activity · last 24 hours peak 31/hr
Summary, timeline and people extracted by Claude from 65 items across 8 sources · 41m ago. Quotes are verbatim.
Anthropic confirmed via an updated Claude support page that it will embed imperceptible watermarks in AI-generated text and add C2PA provenance metadata to generated files, to comply with the EU AI Act's Transparency Code that took effect August 2, 2026. The watermarking, which Anthropic says persists through copying and some editing, will apply to new Claude models globally across products including the API, Claude Code, Cowork and Tag, joining similar moves by Suno, Substack and other AI-adjacent platforms facing regulatory and public pressure over undetectable AI content.
- Anthropic will embed imperceptible watermarks in Claude-generated text and C2PA provenance metadata in generated files, to comply with the EU AI Act's Transparency Code (effective August 2, 2026).
- The text watermark is applied at the model level, persists through copy-paste and some editing, and works across Claude Platform API, Claude, Claude Code, Cowork, and Tag, as well as via AWS, Google Cloud, and Microsoft Foundry.
- New models launched after August 2 get watermarking immediately; support for older models is still being rolled out, and Anthropic has not detailed a public detection tool yet.
- Reactions are mixed: some praise the transparency move, others question its robustness against stripping/editing, and some warn it could mislabel human-written text that was merely AI-copyedited.
How it unfolded
-
Reaction Skepticism about real-world effectiveness
Commenters question whether watermarking will matter in practice, citing lazy users who don't bother stripping AI markers, alongside broader doubts about circumvention.
“I still regularly see chatbot prompts left in emails or student assignments. Most people are very lazy.”
thelincoln.bsky.social · Bluesky ↗ -
Report The Verge details technical mechanics
The Verge explains the two marking techniques—text watermarking and C2PA file metadata—and notes the changes are a future commitment with a four-month grace period for existing models.
-
Reaction Social critics flag copy-edit false positives
Commentators on X note the policy may tag human-written text that Claude merely copy-edited as AI-generated, calling this outcome problematic.
“if I upload text *I* write and ask Claude to copy edit, Claude will now watermark my (human written, AI-copy edited) text as AI, which seems ridiculous”
@petereharrell · X ↗ -
Report Business Insider highlights novel-writing implications
Business Insider reports the watermarking could expose people trying to pass off AI-written novels as human-written.
-
Report The Register covers the pledge
The Register reports Anthropic's commitment to embed watermarks to help trace AI-generated 'slop', framing it as a response to EU pressure.
-
Event Anthropic quietly updates support page
Anthropic adds a new Claude support page detailing how it will comply with Article 50 of the EU AI Act via machine-readable marking of generated content.
- 8 days quiet
-
Event EU AI Act Transparency Code takes effect
The EU's Transparency Code, requiring AI companies to mark AI-generated or edited content detectably, comes into force, triggering compliance moves from major AI vendors.
What people are saying verbatim
“Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.”
Anthropic, Company support page · TechCrunch / The Verge ↗ · Aug 10
“Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported.”
Anthropic, Company support page · The Verge ↗ · Aug 10
“if I upload text *I* write and ask Claude to copy edit, Claude will now watermark my (human written, AI-copy edited) text as AI, which seems ridiculous”
@petereharrell, X user · X ↗ · Aug 10
“The era of unprovable AI writing just ended. Starting August 2, every new Claude model weaves an invisible watermark into the words it generates.”
@aakashgupta, X user · X ↗ · Aug 10
“The modern equivalent of cartographers putting in fake towns and roads on their maps to catch people that copy them.”
karma_dumpster, Reddit commenter · Reddit ↗ · Aug 10
“This just makes AI less useful because it’s more worried about maintaining a hash than actually providing the best result.”
No_Reindeer8688, Reddit commenter · Reddit ↗ · Aug 10
“I still regularly see chatbot prompts left in emails or student assignments. Most people are very lazy.”
thelincoln.bsky.social, Bluesky user · Bluesky ↗ · Aug 10
Voices from the web unedited
-
If LLM watermarks became industry standard, it'd be very interesting. I’ve seen people saying scammers would retype the document but… I dunno. I still regularly see chatbot prompts left in emails or student assignments. Most people are very lazy.
-
They address this in the presentation how if the hash function is indistinguishable from random (and CSPRNGs like say SHA-256 are if you don't know the key) and the candidates you use in your tournaments are the top n ranked candidates with equal scores, then the distribution of the LLM + SynthID ends up matching the distribution of the LLM alone…
-
Anthropic added a new support page today with details about how they would comply with Article 50 of the EU AI Act. As a result, models released in the EU will support machine-readable marking at launch, including text generated via APIs and through providers like AWS, Microsoft and Google…
-
Anthropic says new Claude models will embed invisible watermarks in all generated text, everywhere Claude is offered. The watermark is part of the text, it isn't metadata: “it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing.” [image]
-
this is interesting; presumably apart from EU compliance this is also about preventing model collapse in the longer-term, so they *should* be incentivized to do it—and for book-length texts, it certainly ought to work
-
It does affect the text then? One couls say that it doesn't matter since the chosen output would have been random, but random is a choice nonetheless, an evolutionary mutation in darwinistic terms. And as we know in evolutionary systems, random brings a much better output than a chosen one. You what I mean?
-
This really isn't a great solution. The article explains why. https://www. businessinsider.com/anthropic- watermarking-feature-stops-undetected-ai-generated-writing-2026-8 # ai # watermark # writing
-
If I read Anthropic's new watermarking policy correctly, not only will Claude watermark text it writes as AI (fair enough), if I upload text *I* write and ask Claude to copy edit, Claude will now watermark my (human written, AI-copy edited) text as AI, which seems ridiculous... [image]
-
Wow. At last, Anthropic is going to add watermarking to Claude AI-generated text. Sure, people will find ways to evade it, but at least there'll be a norm of transparency, and we'll get to see whether watermarking works in the real world. — Thank you, EU AI Act! — support.claude.com/en/articles/ ...
-
This just makes AI less useful because it’s more worried about maintaining a hash than actually providing the best result.