US artificial intelligence company Anthropic has announced the deployment of an invisible, tamper-resistant watermark in text generated by new versions of Claude, its AI assistant, enabling authorised tools to identify content that has been processed by the company’s AI models.
The technology was developed to embed a digital signature directly into the token structure of generated text. According to Anthropic, the watermark is imperceptible to users and does not alter the meaning, quality, or style of content produced by Claude.
Unlike traditional metadata, which can be stripped when text is copied and pasted, the new watermark travels with the content even when it is transferred to other applications. The company also states that the system was designed to resist superficial modifications to the text.
The measure is linked to commitments made by Anthropic under the Transparency Code of Practice established by Article 50 of the European Union’s AI Act. Although the requirement is tied to European legislation, the technology has been applied globally across all new models released from 2 August onwards.
The watermark is present across different modes of Claude deployment, including Anthropic’s own application, the API for developers, and enterprise tools such as Claude Code, Cowork, and Tag. The technology is also available on partner cloud computing platforms.
For visual content — including PNG, JPG, and SVG files — Anthropic has adopted the open C2PA standard, which allows cryptographic metadata to be added to indicate the origin and provenance of a file.
Despite the technology, the company clarifies that the presence of a watermark does not necessarily mean the underlying idea was created by artificial intelligence. For instance, a user may write a text independently and use Claude solely to correct grammatical errors, translate, or rephrase the content — in which case the text may also receive the signature.
Conversely, extensive rewrites, broad paraphrasing, or certain translation processes can eliminate the signal and render the watermark undetectable by analysis tools.
Claude models released before 2 August do not yet carry the active feature, but Anthropic has indicated a gradual rollout during the transition period defined by European legislation. The company is also expected to publish technical guidance and tools to allow third parties to verify the presence of watermarks in texts and files.
The initiative places Anthropic among technology companies seeking to increase transparency around content produced or processed by artificial intelligence, while simultaneously raising new questions about authorship, privacy, and the traceability of AI-generated text.
Source: Diário Económico
Original article: https://www.diarioeconomico.co.mz/?p=529626
