Claude content will soon get invisible AI watermarks
Samuel Boivin/Getty Images
Anthropic’s models are about to start telling on themselves. Yesterday, the company announced plans to watermark anything that Claude creates or edits, in compliance with European transparency regulations.
Everything the light touches. The policy will apply worldwide to all text and files that Claude models spit out, whether they’re AI-generated cover letters or AI-proofread lines of code. Loopholes exist: Heavy editing or file conversions may shed the watermark. Otherwise, the virtual scarlet letter will be invisible to the human eye but detectable by Nicolas Cage squeezing lemon juice on it machines. Anthropic said it will help users and other third parties detect the mark, with details to come in “forthcoming documentation.”
Timeline: Watermarking will come with any Claude model launching after Aug. 2, which is when the EU AI Act’s Transparency Code kicked in. It carries a four-month grace period for existing models, so Anthropic said it’s working to get its older Claudes up to par.
Other companies, including Google, Meta, Microsoft, and OpenAI, have also pledged to follow the EU’s new rules.
Zoom out: Anthropic’s move comes as backlash mounts against sneaky AI content. Allegations of computer-generated writing plague the book-publishing world. Meanwhile, AI music platform Suno recently said it would watermark its tracks, and Substack recently launched a tool for flagging AI writing that postures as human-made—something the newsletter company’s CEO referred to as “Claudefishing.”
No comments:
Post a Comment