AI
How Anthropic's AI-Generated Text Watermarks Will Work—and Why They're Hard to Remove
Anthropic has detailed its plan to embed invisible watermarks in Claude's output to meet EU AI transparency rules—and what happens when users try to hide them.
TechCrunch — AI
Anthropic has released fresh technical details about how it will embed invisible watermarks into text generated by Claude, addressing user concerns and technical questions that emerged following its announcement earlier this week.
The Watermarking Mechanism#
The system works by having Claude make strategic choices between semantically equivalent words—such as "overcast" versus "grey" to describe weather—while inserting a subtle pattern into its responses. This pattern remains imperceptible to human readers but becomes detectable to anyone possessing the decryption key, allowing AI-generated material to be identified without compromising readability.
Anthropic emphasized that the technique produces no quality degradation: watermarked output reads identically to unwatermarked text from a user's perspective. The approach draws from SynthID-Text, a methodology developed by Google DeepMind in 2024, and the company plans to release a detection API for verification.
The Compliance Picture#
The watermarking requirement stems from the EU AI Act's Transparency Code, which mandates that AI developers implement systems capable of identifying machine-generated content. The announcement has sparked considerable backlash, with some users canceling subscriptions and framing the measure as surveillance. However, Anthropic notes that other major model developers have adopted similar commitments under the same Code of Practice.
Removing or Bypassing the Watermark#
The watermark proves resilient to casual editing: light revisions typically fail to eliminate it. However, a complete rewrite—where every word is replaced—will effectively destroy the watermark, though Anthropic observes that such extensive rewriting raises the question of whether the output can meaningfully be called AI-generated anymore.
For text that has been merely proofread or lightly edited by Claude, watermark detectability hinges on text length and editing intensity. When human-authored content dominates and Claude's contribution is minimal, "there's very little (if anything) for the watermark to attach to."
