Anthropic will watermark Claude's text output

Anthropic explained on August 14, 2026 that future Claude models will embed a watermark in generated text, to comply with the EU AI Act. It works by steering the model's choice among equally viable next words using a key plus the preceding words, leaving a pattern readable only with a decryption key. It carries no identifying information and cannot be traced to a user, organization, or conversation, and has minimal effect on code or on heavily edited text. Users cannot opt out; a detection API is planned.

Why it matters

Provenance marking moves from proposal to shipped behavior at a frontier lab, and the no-opt-out design means anyone publishing Claude-written prose should expect it to be detectable by whoever holds the key.

Who should care

Anyone publishing model-generated text, and teams with AI-disclosure obligations

What you can do

Nothing to configure — it applies to future models automatically. Watch for the detection API if you need to verify provenance yourself.