On August 2, 2026, Anthropic began embedding invisible watermarks into new versions of Claude—machine-readable labels that remain within the statistical distribution of the text and persist through copying and partial editing. This is not merely a technical update, but a strategic shift from attempting to detect AI content post-factum to proactively embedding proof of origin at the generation stage.
The mechanism operates at the token level: the model adjusts word choice probabilities so that a statistically detectable pattern emerges in the sequence, without affecting meaning, quality, or readability. The company promises verification tools, but the algorithm's details and resilience against attacks (translation, paraphrasing, mixing with human text) have not yet been fully disclosed. Unlike previous approaches based on style and frequency analysis, Anthropic's watermark is an active signal embedded by the manufacturer.
Anthropic's methodology relies on the requirements of Article 50 of the EU AI Act, which came into force on August 2. The regulator requires generative AI systems to add machine-readable labels to identify AI content. This explains the synchronization with other players: Google is expanding SynthID to text, OpenAI is promoting C2PA and digital watermarks, and Meta is testing similar solutions. However, Anthropic stands out by linking the watermark directly to compliance, rather than just internal safety initiatives.
Compared to traditional AI detectors (which analyze perplexity and burstiness), a watermark provides a more reliable signal about a specific model, but it does not solve the problem of mixed authorship. A study published on August 8 on arXiv emphasizes that in the era of human-AI collaboration, the binary division of "human or AI" is obsolete. The watermark records the fact that Claude was used, but it does not show what proportion of the text belongs to a human, what edits were made, or who bears final responsibility.
The limitations are obvious. Large-scale editing, translation, or mixing with other content can destroy the signal. The absence of a watermark does not prove human authorship—it only means that the label was either not embedded or has been removed. Thus, the tool is useful for platforms and publishers as an additional signal, but not as "irrefutable proof."
For the industry, this signifies a transition to a provenance ecosystem: in the future, models will not just generate text but also leave cryptographically verifiable traces. This will simplify moderation but complicate questions of authorship, citation, and legal liability. Researchers have yet to answer how watermarks from different providers will interact and whether a universal detection standard can be created.
Ultimately, Anthropic's watermarks demonstrate that regulation is already defining the technical agenda: content traceability is becoming a mandatory feature of models, rather than an optional function.
