Anthropic will watermark Claude's text with SynthID to meet the EU AI Act

Future Claude models will carry a SynthID-Text watermark that biases word choice with a secret key instead of adding tokens — no extra serving cost, and raters comparing watermarked and plain answers saw no quality gap. The method is real but narrow: detection fades on short samples, tightly factual passages, light proofreading, and code, and a full rewrite strips it. A detection API is coming, and other labs signed the same EU Code of Practice, so provenance is settling into per-vendor keys over long AI prose rather than one universal test.

Source: anthropic.com ↗

Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude's responses.

Anthropic

Why this matters

→ EU AI Act compliance now baked into Claude's text at scale
→ Detection fades on edits, code, and factual passages—not universal proof
→ Per-vendor watermarks settling industry standard over universal detection

Fingerprinting text, not tokens