Detect AI Text: Build a SynthID Watermark Scorer

On August 14, Anthropic published the details of how Claude’s new text watermark works. Every new Claude model launched on or after August 2, 2026 embeds a machine-readable mark in the text it writes — not hidden characters, not extra tokens, nothing you can see. It is a statistical fingerprint left in the model’s word choices, and it is there to satisfy the EU AI Act Code of Practice on Transparency of AI-Generated Content.

The method is a variant of SynthID-Text, published by Google DeepMind in Nature in 2024. Anthropic says a detection API is coming but has not shipped it. That leaves a gap: people are arguing loudly about what this watermark can and cannot catch, and almost nobody arguing has actually measured one.