Text will carry a hidden "statistical pattern," but image pixel watermarking is notably absent.
When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works .
The legislation is accompanied by a Code of Practice with suggestions for implementation, and in keeping with that code, all new Claude models will watermark text and add provenance data to generated files, namely images (of SVG, PNG, and JPG file types). Anthropic's guidance is a bit late to the party, as OpenAI and Google already published their own versions a while back.
When it comes to text, unlike steganography in images that hides data in the picture content, the lab says that watermarking is performed by biasing the selection of tokens (parts of words) during generation. When the generated text is analyzed, it'll fit a determined statistical pattern, revealing the watermark. Anthropic says it'll provide detection tools that look for these patterns in "forthcoming technical documentation."
Text that is copied-and-pasted and only lightly edited should retain the identifiable pattern. Anthropic says that the quality and meaning of the generated text won't suffer as a result of the marking process. Slices of text under 200 tokens are exempt under the Code of Practice, as they don't carry sufficient data to reliably watermark.
As for images, the aforementioned file types support additional metadata attached to the picture itself, and Claude will start adding a provenance certificate using the C2PA standard. In simplified terms, the files will carry an associated digital certificate that will be invalidated if the file is altered in any way.
Alibaba bans Anthropic's Claude Code after an alleged hidden China-detection backdoor is uncovered
Anthropic warns Claude AI is building itself faster than expected, calls for option to halt frontier development
Attentive readers might note there's no mention of actual image watermarking, and indeed Anthropic made no mention of that feature, although it's a requirement of the Code of Practice for images, alongside the provenance information. It's expected the firm will implement image watermarks at some point, otherwise, just copy-pasting the picture content would make it untraceable.
Anthropic also says that it's working to add these capabilities to existing Claude models — as required by the AIA, with a deadline of December 2, 2026. The company's guidance starts by describing "models launched in the EU," but subsequent paragraphs clarify that "marking will apply to output from supported models wherever Claude is offered, worldwide."
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/tech-industry/artificial-intelligence/SPONSORED_LINK_URL
- https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-will-begin-digitally-watermarking-marking-ai-generated-text-and-images-anthropic-details-how-itll-comply-with-the-eus-artificial-intelligence-act#main
- https://www.tomshardware.com/membership
- QIDI Plus5 3D printer review: The best one yet
- Nvidia reportedly testing lower memory configs of Rubin Ultra as memory shortage bites back — designs tested include as little as 192 GB and step back to HBM4
- Nova Lake CPUs with cut-down E-core clusters may still retain full cache pool, says new leak — 8P+12E config predictions revised from 33MB to 36MB, 4P+4E config
- Powerful Compute So Compact, It’s Clutch — Build AI Anywhere With NVIDIA Jetson
- GeForce RTX 50-series GPU prices spike as much as 39% as Blackwell price hikes hit the US — RTX 5070 gets a 36% hike, RTX 5060 up 27% at the median of Newegg li
Informational only. No financial advice. Do your own research.