Overview
- Anthropic said models released on or after August 2, 2026 will carry invisible machine‑readable marks and the company has a grace period until December 2, 2026 to add marks to older Claude models.
- The marks will appear inside text as subtle, aggregate signals and as C2PA provenance metadata for images, audio, and video and Anthropic plans to provide tools to help detect them.
- Anthropic will mark content the models 'process,' including edits, proofreading, translations, and summaries, so the watermark can signal AI involvement even when a human wrote the original text.
- Researchers and reporters warn the approach is fragile and reversible because editing, re‑saving, screenshots, format conversion, or simple metadata changes can remove signals and public detection details could enable circumvention.
- The step reflects EU enforcement pressure and broader industry moves such as DeepMind’s SynthID and carries legal stakes under the EU AI Act where violations can draw fines up to €15 million or 3% of global turnover.