Overview
- Anthropic began embedding a hidden, token-level watermark in Claude models released since August 2 and says it will add marks to older models and publish a detection API that is not yet available.
- The watermark works by biasing Claude’s token choices so text carries a statistical pattern rather than a visible label, and Anthropic says the mark does not identify individual users.
- Because the mark is applied during token selection, Claude will mark text it only edited or proofread, creating a risk that human-written work can test positive as AI-processed.
- The technique is fragile: short passages, code, translations, heavy edits or paraphrasing can erase or weaken the signal, and open-source developers have already released tools that attempt to remove or obfuscate the watermark.
- The global rollout, lack of public detectors or error-rate data, and the EU’s resilience guidance leave open legal and editorial questions about verification control, false attribution, and a likely cycle of marking, detection and circumvention.