On August 14, 2026, Anthropic announced that future Claude models will generate watermarked text so a detector can estimate whether Claude was involved in producing a passage. Anthropic says the change is intended to comply with the EU AI Act's requirements for marking AI-generated content and will launch globally. The important detail is that the mark is not visible to readers. It is a statistical pattern that can be checked with the relevant key.
The method uses low-stakes choices made while selecting the next word. When two words are both reasonable, the watermarking scheme uses a key and a small amount of preceding context to choose the source of randomness. A sequence of these choices can then be tested for consistency with the key. Nothing is added to the text, there are no hidden characters, and readers do not see a label. Detection depends on the key and a sufficiently long sample.
Anthropic says watermarking will not practically change the quality, content, or readability of Claude's outputs, citing internal testing and Google DeepMind's SynthID-Text research. Those are vendor and related-research claims, not an independent guarantee for every model, language, and workflow. Organizations should test false positives and false negatives across their own model versions, prompts, languages, and output lengths.
The limits are explicit. A watermark can answer only how likely it is that Claude wrote or processed the text. It cannot prove that a passage was written by AI or rule out another model. Short samples provide less statistical evidence. If Claude only proofreads a small amount of human text, there may be too few words of Claude's own for a detector to register.
Code is a special case. When an output has to be exact, such as an arithmetic result or syntax that must not change, there may be no equally good alternative choice, so the watermark does not nudge the selection. Anthropic expects code to carry less watermarking than ordinary prose, although comments and other flexible text can still contain a pattern. That makes the watermark a provenance signal, not a complete content classifier.
For users, Anthropic says watermarking needs no extra tokens, has negligible effects on speed and cost, and contains no personal, organizational, or conversation identity. Anthropic plans to offer a watermark detection API later. For supported images, documents, and SVGs, Claude will attach a C2PA content credential in metadata saying that Claude was involved in producing or processing the file. Both signals describe provenance; neither proves authorship or assigns legal responsibility.
The broader shift is from relying on a detector's stylistic guess toward using a model provider's verifiable generation signal. A practical governance design should treat watermark detection as one part of provenance, alongside human review, version history, source records, permissions, and retention rules. Anthropic's method still has limits around short text, rewriting, translation, and content passed between models. It is best used as additional evidence, not as a standalone basis for blocking, attribution, or punishment.
