Anthropic Adds Invisible Watermarks to Claude AI Text

1794089213 Claude


Anthropic is adding an invisible watermark to all text generated by its Claude AI. Every model released since early August already has this hidden tracking feature, and the company is updating older systems before a strict European Union regulatory deadline in December.

This week, outputs from Claude Mythos 5, Opus 4.7, and Sonnet 4.6 will include the machine-readable mark. Anthropic delayed the rollout for Opus 4.6 until November 16, when older models like Opus 4.5, Sonnet 4.5, and Haiku 4.5 will also get the update.

Since the company adds this tracking at the core model level, users cannot turn it off. The watermark shows up no matter how you use Claude, including the official web apps, API connections, and third-party cloud hosts like Amazon Web Services and Google Cloud.

To do this without readers noticing, Anthropic uses a modified version of Google’s SynthID technology. The system slightly changes Claude’s word choices to create a mathematical pattern hidden in the text.

According to internal documents, this invisible signature survives basic copying and pasting, but heavy editing or translation can break the pattern. The company stressed that the watermark does not contain any personal user data or conversation history.

Currently, regular users cannot verify these text watermarks. Anthropic has limited its detection tool to a private preview for law enforcement, researchers, and media organizations.

There are some technical loopholes in the tracking system. For developers using the API, the watermark only works reliably on older models if settings like temperature and top_p stay at their default values. Changing these settings can bypass the tracking. Also, the system cannot tell the difference between text Claude wrote from scratch and a human-written document that the AI only checked for grammar.

Anthropic made this global change to comply with Article 50 of the new EU AI Act. This worldwide rollout is very different from OpenAI, which only enforces watermarks for European ChatGPT users and makes the feature optional for its API customers.

 



Source link

متعلقہ پوسٹ