Anthropic shares more details about how Claude’s new watermarks will work

Following recent announcements regarding its commitment to the EU AI Act’s Transparency Code, Anthropic has moved to clarify the mechanics behind its upcoming watermarking system for Claude. As the industry grapples with the implications of identifying AI-generated content, the company is addressing user concerns regarding privacy, output quality, and the technical feasibility of its new security measures.

Addressing the Transparency Mandate

The decision to implement watermarking stems from the EU AI Act, which mandates that AI developers provide mechanisms to distinguish machine-generated content from human-authored text. While the move has sparked a spirited debate across platforms like Reddit and X—with some users labeling the initiative a "conspiracy" and others threatening to cancel subscriptions—Anthropic is pushing back against the narrative that these measures are intended to track or police individual users.

Instead, the company emphasizes that the technology is a technical response to regulatory requirements, designed to be invisible to the average reader.

How the Watermarking Mechanism Functions

Anthropic’s approach relies on subtle, statistical patterns embedded within the model's output. When Claude makes "low-stakes" linguistic choices—such as selecting between synonyms like "grey" or "overcast"—it can subtly influence the response structure.

• Undetectable to Humans: The watermark does not alter the quality, tone, or readability of the text.
• Key-Based Detection: While invisible to the naked eye, these patterns are easily identified by anyone possessing the specific cryptographic key used to encode them.
• Industry Standards: Anthropic confirmed it is utilizing the SynthID-Text framework, a methodology originally developed by Google DeepMind in 2024.

"Watermarking does not impact the quality of Claude’s output. To a reader, a watermarked response is indistinguishable from an unwatermarked one."

Distinguishing Watermarks from AI Detectors

It is important to note that this watermarking system is fundamentally different from third-party AI detection tools. Companies like Pangram often rely on identifying specific linguistic "tells" or stylistic patterns common in LLM outputs. Anthropic clarifies that its watermark is a deliberate, encoded signal rather than a heuristic analysis of writing style. Furthermore, the company plans to release a dedicated watermark detection API to facilitate transparency.

Can the Watermark Be Removed?

A primary concern for users is whether the watermark can be bypassed through editing. Anthropic admits that while the system is robust, it is not invincible.

• Light Editing: Minor tweaks or proofreading are unlikely to strip the watermark, as the underlying statistical pattern remains largely intact.
• Complete Rewrites: If a user performs a total rewrite where every word is replaced, the watermark will be lost. However, the company notes that in such cases, the text can hardly be considered AI-generated in any meaningful sense.
• Human-AI Collaboration: When Claude is used only to edit human-written text, the watermark's presence depends on the extent of the model's involvement. If the human author retains the bulk of the writing, there is insufficient AI-generated "signal" for the watermark to attach to.

The Impact on Code

The technical constraints of programming languages provide a natural limit to watermarking. Because code must be syntactically correct to function, the model has less flexibility to make arbitrary word choices.

"Code should have less of a watermark than other text," the company explained. "In areas where there is an arbitrary choice between particular words or terms—such as comments within code—the watermark can be used. But by definition, it will have a negligible effect on the actual code produced."

As the AI landscape evolves, Anthropic expects this to become the industry standard. The company noted that other major developers have signed the same Code of Practice and are expected to roll out their own proprietary watermarking solutions in the near future.