Anthropic plans to watermark text from future Claude models using a statistical pattern in word choices. The company presents the system as a transparency measure for the EU AI Act and says it does not add hidden characters, user identifiers, or practical quality loss.
Transparency is a legitimate goal. A probabilistic watermark becomes dangerous when a school, employer, publisher, or compliance team reads it as proof.
Probability is not authorship
Anthropic says its detector will estimate the probability that Claude contributed to a text. It will not prove who wrote it, why Claude was used, whether a person reviewed it, or which parts remain from the original output.
The company also identifies weaker detection for short text, factual answers, code, and lightly edited material. Heavy rewriting can remove the signal, while translation may preserve it. Those limits make the same score mean different things across workflows.
A global watermark exports one region's compliance model
Anthropic says it will roll out watermarking globally because durable regional separation is difficult. Users outside the EU therefore inherit a control designed around one regulatory requirement.
That may simplify the provider's architecture, but it shifts the policy choice to every user and organization. A company may receive watermarked text without having chosen the detector, the standard, or the interpretation downstream platforms apply.
Governance needs evidence from the workflow
A detector result can support an investigation. It should not decide academic misconduct, employee performance, publication status, or contractual compliance by itself.
Organizations need stronger evidence: model-use logs, version history, declared assistance, review records, source checks, and ownership of the final decision. Those controls explain what happened. A watermark only suggests that one model may have participated.
Watermarking can add a provenance signal. Opposing it as a primary compliance mechanism is reasonable because probabilistic detection can create false certainty precisely where the consequences require proof.