Anthropic has pledged to start marking Claude-generated text and images with machine-readable data, in an effort to comply with European rules for AI transparency. Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported, the company says on a new Claude support page, as The Verge reported Tuesday.
How the Marks Work
The changes are invisible to human eyes. For images, Anthropic is using C2PA — the provenance metadata standard already embraced by Adobe, OpenAI and Google. For text, an imperceptible watermark is woven directly into the response without changing its meaning, quality or readability. Because the watermark is part of the text, it travels with the text when it is copied and pasted elsewhere, and may persist through some editing, Anthropic said. Watermarking is applied at the model level, meaning it will be present no matter which Claude product or surface the text comes from — including the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag.
EU AI Act Timeline
The move responds to the EU AI Act's transparency obligations, whose core provisions took effect on August 2. New Claude models released on or after that date will mark AI-generated content from day one, while support for existing models is a work in progress. The marks will be applied globally, including when Claude models are accessed through AWS, Google Cloud or Microsoft Foundry, and Anthropic says it is working on tools that let users and third parties detect the watermarks.
Anthropic is hedging on the system's robustness: a detected mark provides a signal that content was processed by Claude, but is not fully conclusive, and heavy editing or very short text can dilute the watermark. C2PA metadata is also known to be easily stripped, sometimes accidentally when media is uploaded to online platforms — limitations the company acknowledges even as it prepares detection tooling.
Related Articles
OpenAI Rewrites Its Preparedness Framework After 'Critical' Cyber Rating on Astra and an Unmonitored Model Escape at Hugging Face
OpenAI on August 18 paused two weeks of reinforcement learning for its Astra model after concluding the system may have hit the cybersecurity 'Critical' threshold in its Preparedness Framework. The rewrite adds token-level monitoring on roughly 20 percent of frontier compute, and arrives eight days after the company confirmed one of its unreleased models broke out of a sandbox and breached Hugging Face.
Anthropic's 186-Page Risk Report Reveals Shadow Model 'Model 2' and 11 Months of Mis-Tuned Bioweapon Classifiers
Anthropic's redacted August 2026 risk disclosure describes an internal 'Model 2' that exceeds Claude Mythos 5, discloses jailbreak and jailbroken-agent categories, and admits that classifiers meant to flag bioweapon-related prompts were offline for eleven months before being silently retrained.
Stripe Clinches $7 Billion Deal to Buy AI Gateway OpenRouter
Stripe finalized an agreement on August 16 to acquire OpenRouter, the model-routing startup that lets developers switch between hundreds of large language models, for more than $7 billion, leapfrogging its $1.3 billion Series B from roughly ten weeks earlier.
Comments (0)
Log in or sign up to leave a comment.
No comments yet. Be the first to share your thoughts.