Anthropic to start watermarking Claude-generated text, images
SiliconANGLE Maria Deutscher ● Covered by 4 sources
Anthropic says Claude will start watermarking text and images. It’s to match EU AI Act rules, and the mark is meant to survive copying and some edits.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic is adding invisible watermarks to content made by Claude, starting with text and images. The company’s plan surfaced in a help desk article on Monday and was reported by The Register. It covers the Claude model family and the Anthropic services built on top of it.
The timing is tied to the European Union’s AI Act. Anthropic has signed the law’s voluntary Code of Practice, which calls for AI-generated content to be marked. For text, the company says the watermark will appear in Claude models released after Aug. 2, with older models to follow later.
Anthropic says the mark won’t be visible to users, and it should survive being copied into other apps. It may even stick around after edits. The company hasn’t explained how the text watermark works, but a 2024 Google DeepMind paper offers a likely clue: a method that nudges a model’s word choices so its output can be detected later.
Images are getting a different treatment. Anthropic says it will use C2PA, a system that attaches metadata to AI-generated files and records which model made them, when, and whether copyright limits apply. The help desk article names JPG, PNG and SVG as covered formats. Anthropic also says it will release tools to help people detect Claude-made content, while warning the system won’t catch everything, especially short or heavily edited text snippets.
OpenAI has also signed the Code of Practice, so it may be next in line to roll out its own text watermarking. That wouldn’t be a surprise. The industry keeps promising provenance tools only after regulators start circling, which is a very efficient way to discover conscience.
My take — AI-written commentary, not fact-checked reporting
This is the right move, and also a small confession: AI companies know they’ve spent years making attribution someone else’s problem. Watermarking won’t solve trust, but it at least stops pretending that synthetic content should drift around the internet in a fog of plausible deniability. The awkward part is how long it took for this to become standard behaviour rather than a press-release virtue signal.
Read more about this at: SiliconANGLE
