How Claude's text watermark works

What changed

Source summary

Anthropic is adding an invisible text watermark to future Claude models, using a SynthID-Text-style method that changes word-selection randomness without affecting output quality, to comply with the EU AI Act. The watermark can't identify individual users or organisations and doesn't slow performance or increase costs, but it has real limits: it can't reliably distinguish Claude-written text from heavily-edited text, works poorly on short or highly factual passages with limited word choice, and can be removed by a full rewrite.

Why it matters to your work

Practical interpretation

Anthropic is adding an invisible watermark to Claude's text output, starting globally — not just for EU users. It switches on automatically with no setup. Key caveat: it can't reliably prove whether a document was AI-written versus human-edited, works poorly on short or highly factual text, and a full rewrite removes it — so it's not an academic-integrity or authorship-proof tool.

When this could help

Relevant now for organisations evaluating AI content detection (e.g. education, publishing, compliance). No user action needed — watermarking is applied automatically to all Claude output at launch.

Anthropic now watermarks all Claude-generated text using SynthID-Text (Google DeepMind 2024, Aaronson 2022). The watermark changes the randomness source when choosing among similarly good words — key plus preceding words determine the choice. No hidden characters, no extra tokens, no quality/speed/price impact. Detection API is in private preview for regulators, law enforcement, media, fact-checkers, researchers, educational orgs, EU civil society, and obligated enterprises.

Build it into your routine

SCALA suggested workflow

1. AI content policy update

   When reviewing your organisation's AI content policy, note that Claude output is now watermarked and a detection API exists in private preview — this may be relevant for academic integrity, content provenance, or compliance workflows.

What to check

• Watermark is sparse in factual, exact, and code passages where word choice is constrained. Light editing may not remove it, but a full rewrite will. Cannot distinguish 'written by AI' from 'heavily edited by AI'. Source ↗
• No detection accuracy figures are published. No user or organisation tracing — the watermark identifies AI-generated text, not who generated it. Source ↗