Anthropic details Claude text watermark as probabilistic, sparse in code and factual text, and removed by rewrite

Anthropic has publicly described the properties of a text watermark that will be present in future Claude model outputs. The three properties stated in the single feed carrying the story are: the watermark only signals that Claude was likely involved, it is sparse in code and factual text, and it disappears after a full rewrite. No technical mechanism, no model version, and no rollout date appears in the material provided, so any claim about how the watermark is implemented, which model introduces it, or when it ships would be invented rather than grounded.

The 'likely involved' framing is the load-bearing design choice. A probabilistic signal carries an inherent false-positive and false-negative rate, and Anthropic is explicitly putting the watermark in that category rather than presenting it as a deterministic identifier. For an engineer this means any detector built on the watermark must return a confidence value and a threshold, not a binary attribution, and must be prepared to be wrong on both sides of the threshold.

The sparsity in code and factual text is the more consequential caveat. Code generation, technical documentation, and factual reporting are precisely the use cases where questions of authorship, licensing, and AI assistance are most often litigated or audited. A signal that thins out in those domains is a signal that cannot carry the weight an attribution system would want. Any detection pipeline that ingests Claude-generated code or factual prose should expect weak or absent watermark coverage there, and should not depend on the watermark as a gate.

Defeat by full rewrite closes the loop. Any post-processing step that rephrases Claude output while preserving meaning - human editing, a paraphrasing model, a translation round-trip, even a sufficiently aggressive copy-edit - will remove the watermark. That includes the workflows publishers and engineering teams already use to clean up AI-generated drafts, which means the watermark will survive in exactly the workflows where it is least useful (raw, unedited model output) and vanish in the workflows where attribution matters most (final, edited artifacts).

Because only one feed is carrying the event, the description of behaviour is single-sourced and should be read as Anthropic's own characterisation rather than an independently verified property. Engineers planning to depend on the watermark should wait for independent measurement of false-positive rate, sparsity in different content types, and robustness against known rewrite attacks before treating the stated properties as design constraints they can build against.