source: TechCrunch AI: Anthropic shares more details about how Claude’s new watermarks will work
level: business
Anthropic published a blog post explaining how it will watermark text generated by its chatbot Claude. The move follows the company's earlier announcement that it would add watermarks to comply with the EU AI Act's Transparency Code. The code requires AI companies to use systems that make it possible to identify AI-generated content. The announcement sparked debate among users, with some on Reddit calling it a conspiracy and others claiming to cancel subscriptions.
Anthropic said it will use the SynthID-Text approach outlined by Google DeepMind in 2024 and plans to release a watermark detection API. The watermark is created by making low-stakes choices, such as choosing between 'overcast' and 'grey,' to form a pattern detectable only with a key. Light editing probably won't remove the watermark, but a complete rewrite will. Code will have less watermarking because the model must produce working code, though comments can be watermarked.
Anthropic noted that watermarking is distinct from AI detection tools like Pangram that look for writing patterns. The company said other major model developers have signed the same Code of Practice and will implement their own watermarks. This means watermarked text may become common across AI chatbots. For users, the watermark is invisible and does not affect output quality, but it allows platforms and regulators to verify AI-generated content.
why it matters: Watermarking lets platforms and regulators verify AI-generated text, which matters for transparency and compliance with laws like the EU AI Act.
source: TechCrunch AI: Anthropic shares more details about how Claude’s new watermarks will work