Anthropic adds an invisible SynthID-style watermark to Claude’s text

Your AI and this story

Key facts

• Published Aug 14, 2026 (‘How Claude’s text watermark works’); future Claude models generate watermarked text, and models launched before Aug 2, 2026 (covered by the EU transition period) get it over the following months
• Method: among equally valid word choices Claude picks using a cryptographic key seeded by the preceding words, adapted from DeepMind’s SynthID-Text (Nature, 2024), which traces back to Scott Aaronson’s 2022 proposal. Anthropic says it does not change meaning, quality or readability
• Scope: applied globally at launch because Anthropic does not ‘yet have a durable way to scope it by region’
• Detection: Anthropic keeps the key; a detection API is in private preview for groups eligible under EU law (regulators, law enforcement, media, fact-checkers, researchers, educators, civil society) and compliant enterprises
• Files: images and other files get C2PA content credentials, ‘a cryptographically signed note in the file’s metadata’ saying they were made or processed with Claude
• Limits (Anthropic): weak on short samples, code and factual passages with few word choices; full rewrites remove it; it cannot prove that unmarked text is human-written or identify a user
• Follow-up research (Oct 2026, secondary sources): Lasso Security’s ‘The Provenance Tax’ (Andrea Siposova) ran paired tests on seven open-weight models with Hugging Face’s SynthID-Text processor. It reported lower tool-calling accuracy in six of seven and changed refusal behavior under prompt injection, with the effect depending on the watermark key. At temperature 1.0 Phi-4 changed its tool-call verdict on 16.8% of tasks. Lasso did not test Anthropic’s deployed Claude watermark