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

August 15, 2026 | Source: TechCrunch AI

Tags: Anthropic, Claude, EU AI Act, SynthID-Text, watermarking, content provenance

Anthropic adopted Google DeepMind's SynthID-Text watermarking for Claude to comply with the EU AI Act, embedding invisible statistical patterns in word choices—with a public detection API planned, though a full word-for-word rewrite can defeat it.

Details

Anthropic published a technical explainer detailing how it will watermark Claude outputs under the EU AI Act's Transparency Code. Rather than inserting visible markers, the system exploits low-stakes word-choice moments—selecting 'overcast' over 'grey'—to embed a statistically detectable pattern imperceptible to readers. The underlying technique is SynthID-Text, developed by Google DeepMind and published in 2024; Anthropic plans to release a public watermark detection API so third parties can verify Claude-generated content.\n\nThe watermark is resilient to light editing but can be defeated by a complete word-for-word rewrite. Anthropic notes that in the latter case, the text is arguably no longer AI-generated. Code outputs will carry weaker watermarks because programming syntax leaves fewer arbitrary stylistic choices—though code comments remain watermarkable.\n\nUser pushback was immediate: Business Insider reported dozens of subscription cancellations, while Reddit split between those opposing any tracking and those who view AI-content labeling as a basic transparency requirement. The change applies to Claude users in EU-regulated contexts and will expand as Anthropic rolls out its detection API.