From September 30, 2026, additional Claude models are expected to produce marked texts. The watermark cannot be turned off, but it does not provide definitive proof of origin.
Anthropic is expanding its text watermarking to additional Claude models. Starting on September 30, 2026, Claude Fable 5, Claude Sonnet 5, and Claude Opus 4.8 are expected to generate marked outputs. The marking is applied at the model level and cannot be deactivated by users.
More models are expected to follow
By December 2, affected older Claude models are also expected to be retrofitted. According to heise online, Anthropic applies the watermark globally across all Claude platforms. The background for this is the transparency obligations of the EU AI Act.
The process alters the selection between plausible next text segments. This may create a statistical pattern over longer passages, which can be recognized with an appropriate key.
A hit is not conclusive proof
A detectable watermark merely indicates, according to Anthropic, that Claude was likely involved in a text. This could also mean that Claude translated, summarized, or edited a foreign text.
Additionally, the signal can be weakened or removed through heavy rewriting, paraphrasing, translating, or mixing with other text. For short passages, the statistical material may be insufficient for detection. Conversely, the absence of a watermark does not prove that a text is from a human.
What this means for you
The watermark may provide a clue when categorizing longer texts, but it is not a reliable proof of origin. Additionally, a text cannot be readily checked for marking by oneself: the detector is only available through an API in private preview, according to the source. Access is granted to certain organizations and individuals; there is no publicly accessible detector.
Continue learning on TutKit
- Unleashing Claude AI: Your practical training for creative and professional breakthroughs
- Designing with AI: from prompt to finished design with Claude
Comments
Register and comment.