Anthropic expands Claude’s invisible text watermarks to three more models
Fable 5, Sonnet 5 and Opus 4.8 will begin producing marked text on Sept. 30, while older covered models are due to follow by December.
Anthropic will add invisible text watermarks to Claude Fable 5, Claude Sonnet 5 and Claude Opus 4.8 on Sept. 30, extending a marking system that is already used on several newer Claude models.
Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5 and Claude Opus 5 already carry the marks. The company is also retrofitting covered models released before Aug. 2, with all scheduled to receive marking by Dec. 2.
As covered in August, Anthropic began watermarking newer Claude output as part of its response to EU transparency requirements. Its support documentation links the commitments to its signing of the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content.
The programme is global rather than confined to Europe. It applies to supported output across the Claude Platform, Claude, Claude Code, Claude Cowork and Claude Tag, as well as supported models accessed through AWS, Google Cloud and Microsoft Foundry.
The watermark is an imperceptible statistical pattern embedded directly in text. Anthropic creates it by biasing the model’s selection among plausible next tokens, using its own variant of Google DeepMind’s SynthID Text approach. The company says the technique does not alter a response’s meaning, quality or readability, and users cannot turn it off because it is applied at the model level.
The marking is designed to survive copying and pasting, and can persist through some edits. But heavy editing, paraphrasing, translation or combining the material with other writing can weaken or remove it. A complete rewrite in which every word is replaced will remove the watermark, Anthropic has said.
That limitation matters for how detection results should be read. A detected mark indicates that material may have been generated or processed by Claude, not that Claude was the sole source of its ideas or wording. Claude can mark text it has proofread, translated, summarized or edited, while the absence of a mark does not establish human authorship or rule out AI processing.
Anthropic has a detector, but its detection API remains in private preview for eligible organizations and enterprises with relevant verification duties. There is no freely available public detector; the company says it plans to expand access over time.
The company also attaches C2PA-based Content Credentials, or signed provenance metadata, to supported generated files. That is separate from the embedded statistical watermark used for text.
The expansion comes despite swift efforts to evade the system. WIRED previously reported that developers created removal tools after Anthropic confirmed its global watermarking plan, including code by Guillaume Meyer that uses an unwatermarked language model to generate repeated rewrites with substituted synonyms and slight reorganisation.
Anthropic’s remaining rollout target is Dec. 2, when it says all covered Claude models released before Aug. 2 will have marking support.