The regulatory landscape governing artificial intelligence has officially shifted from toothless voluntary safety pledges to binding statutory mandates. Under the EU AI Act, which took effect for market participants on August 2, major AI developers serving European users must mark synthetic text generated by their systems. To comply with this framework and the shared Code of Practice signed across the industry, Anthropic has confirmed that future Claude models will generate text containing an embedded mathematical watermark.
The Sampling Mechanics of Token Watermarking
Large language models operate by predicting and generating text token by token, repeatedly sampling from a distribution of plausible candidates based on preceding context. In scenarios where multiple candidates carry equivalent semantic weight—such as choosing between "overcast" or "grey" following the phrase "The weather today was cold and..."—standard sampling resolves the selection via an arbitrary pseudo-random number generator. Anthropic’s implementation replaces generic entropy with a deterministic key that conditions token selection directly on the preceding sequence.
"Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick."
Because the substitution operates strictly across high-probability candidates, the system never forces the selection of unnatural outliers or obscure synonyms like "nubilous." The resulting statistical bias remains completely imperceptible to human readers while allowing anyone possessing the verification key to mathematically prove Claude's authorship.
Quality Parity and Zero-Cost Inference
For enterprise architectures and product workflows, this watermarking introduces zero computational overhead and alters neither token consumption nor API pricing. Because the technique relies entirely on pseudorandom token distribution rather than hidden unicode characters or injected metadata payloads, the generated text contains zero user-identifying markers linking output to a specific account, organization, or chat session.
The widespread adoption of token-level watermarking establishes a uniform compliance standard across foundational model providers without degrading model intelligence or inflating compute budgets. As frontier developers align behind the same Code of Practice, synthetic text detection will increasingly rely on clean statistical verification rather than invasive, breakable post-processing filters.