Anthropic disclosed that its Claude AI will now embed an invisible watermark in every piece of text it produces, a step designed to satisfy the European Union’s recently enacted AI transparency rules. Unlike visual markers or hidden characters, the watermark consists of a subtle pattern embedded in the word‑selection process, detectable only by an authorized decoder.

The company explained that large language models generate text one word at a time, choosing from a list of plausible options. Normally the choice is driven by a random number generator, but with watermarking Claude uses a secret key to steer the selection. Anthropic illustrated the concept with a simple example: using the digits of pi as a key, the model might pick the sixth word from the candidate list, then the fifth, third, and fifth again, following the sequence 2‑5‑3‑5 derived from the pi digits. The approach adapts Google DeepMind’s SynthID‑Text method, which was detailed in a Nature paper.

According to Anthropic, the watermark does not degrade the quality of Claude’s output, slow down generation, or increase token consumption. Users will see the same fluent prose they expect, while a separate API will allow developers to submit a text block and a decryption key to verify whether Claude was involved in its creation.

Anthropic cautioned that the system has limits. It can flag text that Claude generated or edited, but it cannot distinguish between a full‑length generation and a light edit. Consequently, even a brief proofread will leave a watermark, though very short passages may evade detection. The company noted that translations inherit the watermark as well, and that only a complete rewrite can fully remove the trace.

Code presents a different challenge. Because programming output often requires exact syntax with little room for alternative word choices, the watermark’s pattern is harder to embed. Anthropic said that code typically receives “generally less watermarking” than prose, as the model may have no selection freedom in many cases.

In addition to text, Anthropic will watermark images generated by Claude. Each image will carry a cryptographically signed note in its metadata stating that it was produced by Claude. The company applied watermarks to all Claude products using models released after Aug. 2 and plans to retrofit older models over the next few months, citing the difficulty of implementing region‑specific solutions.

Anthropic’s rollout arrives amid growing scrutiny of AI‑generated content and its potential misuse. By embedding a covert yet verifiable marker, the firm hopes to give regulators, businesses, and end users a tool for tracing AI‑originated material without compromising the user experience.

Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.