This is part of working with the EU AI Act, other labs are adding similar…
- Source
- x.com
- Date
This is part of working with the EU AI Act, other labs are adding similar watermarking. It’s hard to identify AI-generated text, and this gives people better tools to do that. We’ll also be a shipping text detection API that you can use yourself. https://t.co/OZtG0fNsqc
Context
Anthropic's text watermark is designed to estimate whether Claude contributed to a passage. Its subsequent explanation describes a keyed pattern in the model's word choices, rather than hidden characters or a label attached to the document. A detector holding the key checks whether the text fits that pattern. Anthropic's FAQ says detection works poorly on small samples and on human writing that Claude only lightly proofread, where little exists for the pattern to attach to.
The August 11 post promised a detection API. The FAQ's September 1 update describes a private preview for eligible organizations, with broader access planned. Detecting Claude's involvement therefore remains different from identifying every kind of AI-generated writing.
Checking sign-in…
Loading comments…