Anthropic is committing to embed invisible watermarks in Claude-generated text and digitally signed metadata in images to comply with the EU's AI Act transparency rules, which took effect on August 2nd.
New Claude models will include these marks from day one of release, though existing models are still being updated.
While the watermarks travel with copied text and persist through some editing, the company acknowledges the marks are not foolproof and plans to share detection methods in future technical documentation.
What happened
Anthropic pledged to embed machine-readable watermarks in Claude-generated text and digitally signed provenance metadata in images, making AI-generated content detectable while remaining invisible to human readers. New Claude models will carry these marks from launch, while support for existing models is in progress.
Why it matters
The EU's AI Act, which took effect on August 2nd, requires transparency about AI-generated content. Anthropic's watermarking approach—applied globally to Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag—helps the company comply with a four-month grace period for existing AI products. For readers and platforms, the marks offer a way to identify AI-generated text and images without manually analyzing them.
What to watch
Anthropic plans to share technical documentation on how third parties can detect the watermarks and metadata, though it remains unclear whether existing detection tools (such as those in Google's Gemini chatbot for C2PA metadata) will work with Claude files. The robustness of text watermarks is unproven—C2PA image metadata is known to be easily stripped, and Anthropic itself notes that marks are "far from infallible."
Anthropic announced a new initiative to combat the spread of unmarked AI-generated content by embedding invisible watermarks into Claude-generated text and digitally signed provenance metadata into images, as required by the EU's AI Act transparency rules that came into force on August 2nd.
The watermarking will roll out on a staggered timeline. New Claude models will carry these marks automatically from their release date, while existing models—which benefit from a four-month compliance grace period under the AI Act—are being updated as a work in progress. The machine-readable marks will apply globally across all Claude products: Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, as well as when models are accessed through AWS, Google Cloud, or Microsoft Foundry.
Anthropically described the watermarking mechanism in a new support page. For images, the company uses C2PA, a provenance metadata standard that Adobe, OpenAI, and Google have already embraced. For text, Anthropic employs what it describes as an "imperceptible watermark" woven into the generated content without altering meaning, quality, or readability. Crucially, because the watermark is embedded in the text itself, it persists when the text is copied and pasted elsewhere and may survive some editing—a practical advantage over metadata-only approaches. The watermark will be applied at the model level, meaning it appears across all Claude surfaces and products.
Anthropic is also building detection tools to help users and third parties identify the watermarks and metadata, with technical documentation to follow. However, significant limitations remain. C2PA metadata is known to strip away easily, sometimes accidentally, when media is uploaded to online platforms. The robustness of Anthropic's text watermarking is untested. The company itself cautioned that these systems are "far from infallible" and that content lacking detectable marks could still be AI-generated. The article notes that existing detection tools—such as those in Google's Gemini chatbot for C2PA data—may not work with Claude files, though Anthropic said it would clarify this in the future.
Anthropic's watermarking pledge responds directly to the EU's AI Act, which mandated new labeling and transparency obligations effective August 2nd. The company is staggering its rollout: new models will embed watermarks from launch, while existing models receive updates during the four-month compliance grace period. This approach acknowledges both regulatory pressure and practical constraints—updating live systems takes time.
The technical strategy uses two methods: C2PA metadata for images (a standard already adopted by Adobe, OpenAI, and Google, signaling industry alignment) and an undisclosed imperceptible watermark embedded in text itself. Anthropic's choice to build watermarks into the text layer rather than as separate metadata means the marks remain attached even when text is copied and pasted, addressing a real-world detection problem. However, the company explicitly hedges on robustness, noting that marks are "far from infallible" and that content lacking detectable marks could still be AI-generated. This candor reflects the immaturity of watermarking as a detection tool—C2PA image data is known to strip away easily during platform uploads.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Amazon and Google are intensifying competitive efforts against The Trade Desk (TTD), a major digital advertisi…
QumulusAI announced a GPU-as-a-Service agreement with DRW, a global trading firm, to supply a dedicated NVIDIA…
OpenAI introduced Premium Seats for ChatGPT Business, priced at $125 per user per month ($100 with annual bill…
Computer scientists at University of Tübingen, Max Planck Institute, MATS Research, and Snyk discovered a meth…
OpenAI announced that its unreleased model Astra had produced solutions to 10 long-standing mathematics proble…
Anthropic confirmed it will add watermarks to text generated by Claude and other models to comply with the EU…
The AI news that matters, in one minute each morning.
Sign up free