How Claude's Invisible Watermark Works: Anthropic Offers an Explanation

Anthropic Sets the Record Straight on Its Watermark System

According to НВ — Техно: On August 17 at 10:30, Anthropic released an explanation of how the invisible watermark embedded in Claude's text output works. As AI-generated content spreads, hidden markers like these are gaining traction among developers and policymakers. The system is designed to flag AI-written material and aligns with the EU AI Act. Since the mark is imperceptible to readers, it comes from a statistical pattern: during text generation, Claude prefers certain tokens over others based on a secret key held by Anthropic. For example, when describing weather, the model might pick "gray" instead of "gloomy." But a phrase like "Paris is the capital of ..." leaves less room for such choices. Each token selection is invisible on its own, yet across a longer text, a specialized detector can assess whether Claude was the likely source.

Plans and Public Reaction

Anthropic also plans to make a detection tool available through its API. Adding the watermark has almost no effect on generation speed or token consumption. Still, pinpointing the origin of short texts or code is harder, since they feature fewer ambiguous options. The system may even trigger when Claude was only used to polish human-written content. The prominence of the trace depends on the text's length and how heavily it was edited.

The technology builds on the 2024 SynthID research, and Google already uses a comparable labeling system. Some Claude users on X, however, threatened to cancel their paid subscriptions; they don't want employers, teachers, or other parties to know AI was involved. Anthropic told Business Insider that it saw no increase in cancellations after the watermark announcement.

"Invisible watermarking of Claude's output marks a meaningful step toward greater AI transparency." - Business Insider

This move also reflects a growing regulatory emphasis on ethical and responsible AI. The mixed user reaction highlights privacy and control concerns around AI-generated content, and it remains to be seen how this technology will affect AI adoption across various fields.

As the adoption of AI-generated content continues to rise, understanding the implications of such technologies becomes crucial. For a deeper dive into the mechanics behind the watermarking system and its potential impact on content creators, readers can explore our detailed article on invisible watermarks in Claude's outputs.