Anthropic has pledged to commencement marking Claude-generated matter and images pinch machine-readable data, successful an effort to comply pinch European rules for AI transparency. “Generated matter will transportation embedded watermarks, and generated files will see digitally signed provenance metadata wherever supported,” Anthropic says connected a caller Claude support page. The changes are invisible to quality eyes, but will make it easier for group and online platforms to observe if contented was generated by Claude models.
These updates are a early committedness alternatively than thing that will spell into effect immediately. New AI labeling and transparency obligations nether the EU’s AI Act, which came into effect connected August 2nd, see a 4 period compliance grace play for existing AI products that launched anterior to that date. As such, Anthropic says new Claude models will people AI-generated contented from time 1 upon release, but support for its existing models is simply a activity successful progress.
The machine-readable marks will beryllium applied globally to supported Claude models, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. Two different marking techniques are being used: for images processed by Claude, C2PA — a provenance metadata modular already embraced by Adobe, OpenAI, and Google — will beryllium applied to supported files, but the process for marking Claude-generated matter is overmuch lighter connected details.
According to Anthropic, an “imperceptible watermark” is woven straight into the matter generated by Claude models without changing the meaning, quality, aliases readability of the chatbot’s response. Anthropic doesn’t sanction this watermarking system, but says those matter watermarks will besides beryllium applied erstwhile Claude models are accessed done AWS, Google Cloud, aliases Microsoft Foundry.
“Because the watermark is portion of the text, it will recreation pinch the matter erstwhile it’s copied and pasted elsewhere, and whitethorn persist done immoderate editing,” Anthropic says connected the Claude support page. “Watermarking will beryllium applied astatine the exemplary level, which intends it will beryllium coming nary matter which Claude merchandise aliases aboveground the matter comes from.”
Anthropic is besides moving to alteration users and different 3rd parties to observe watermarks and provenance metadata embedded into Claude-generated content, and says it’ll stock specifications connected this discovery strategy successful upcoming method documentation. There are respective devices already disposable that are designed to observe C2PA metadata, including Google’s Gemini chatbot, but it isn’t clear if those will activity pinch Claude-generated files. I’ve asked Anthropic for clarification.
This is different measurement towards AI-generated matter and images being intelligibly tagged crossed online platforms, and a imaginable triumph for folks who want to debar consuming specified content. Fanfiction readers person already been building much rudimentary discovery systems to emblem erstwhile Claude devices person been used successful AO3 fanworks, but these marking systems tin beryllium applied acold much broadly — if they work, that is.
C2PA information is known to beryllium easy stripped out, sometimes moreover accidently erstwhile the media carrying it is uploaded to online platforms, and it’s unclear really robust Anthropic’s matter watermarking solution is. Even the institution itself is hedging that these marking systems are acold from infallible, and that immoderate contented that lacks detectable marks could still originate from generative AI models.
Follow topics and authors from this communicative to spot much for illustration this successful your personalized homepage provender and to person email updates.
English (US) · 
Indonesian (ID) · 
