Claude will start watermarking the text it writes

Anthropic says future Claude models will embed an invisible watermark in the text they generate, so content produced by Claude can later be checked by anyone with the right key. The method is SynthID-Text, first built by Google DeepMind. It works by changing the source of randomness Claude uses when it picks between words that mean roughly the same thing. Instead of choosing at random, the model follows a deterministic pattern derived from a cryptographic key and the preceding text. The wording still reads naturally, but the pattern is detectable to whoever holds the key. Anthropic's write-up compares it to playing Monopoly with digits of pi instead of dice: the game feels the same, but the sequence can be verified afterward.

The company reports no statistically significant difference in how users rated watermarked versus normal output, and no extra tokens, latency, or cost. The watermark also carries no information about the user or the conversation. It has real limits, though. Short texts, heavily factual passages, proofreading edits, and code all give the model few word choices to work with, so the mark can be too sparse to detect.

The change is driven by regulation. Around 190 organizations, Anthropic among them, signed the EU Code of Practice on transparency of AI-generated content in July 2026, which pushes providers to label machine-written text.

Why it matters

If you publish or moderate text, assume Claude's output may become verifiable as AI-written, which helps with provenance but is not a reliable detector for short or code-heavy content. If your process relies on that gap, plan for it to narrow.