Overview
- Anthropic announced detailed plans on August 15 that future Claude models will embed invisible, machine‑readable watermarks and that older models will be retrofitted during the EU transition window.
- The watermark works by biasing low‑stakes token or word choices—a method based on the 2024 SynthID idea—so nothing visible is added to the text and readers cannot tell the difference.
- Detection requires Anthropic’s secret key and a statistical test rather than spotting a visible mark, and the company says watermarking adds no extra tokens and has negligible impact on speed or cost.
- The signal is fragile because it depends on many small word choices, so it can be weak or lost in short answers, exact code outputs, light edits, format changes, or when text is reprocessed by other models.
- Anthropic says it will provide a detector API and keys but has not published detectors, detection thresholds, or empirical error rates, raising concerns about misattribution, centralized verification power, and likely efforts to evade the marks.