Claude's invisible text watermark: what it can prove at work

Anthropic says future Claude models will place an imperceptible statistical watermark in generated text, with a detector API planned. The change is designed for EU transparency compliance, but it will affect policies far beyond Europe because Anthropic plans to apply it globally at launch. The critical limit is simple: a positive result can indicate that Claude was involved, but it cannot establish authorship, intent or misconduct.^{1 2}

Anthropic's approach is based on SynthID‑Text. When several next‑word choices are similarly valid, a secret key influences the random selection, leaving a pattern across a long passage. Nothing visible is inserted, and there are no hidden characters to strip. The signal is statistical, so confidence tends to improve with longer samples.¹

The watermark has less room to operate when only one answer is correct, when code syntax is constrained or when Claude makes a few proofreading edits. Anthropic explicitly says detection does not work well on small samples. A policy that treats every short positive or negative result as definitive would overstate the technology.¹

Anthropic says its detector will estimate the likelihood that a passage was partly written by Claude. It cannot distinguish a Claude‑written draft from a human draft that Claude heavily edited, and it cannot identify a user, organization or chat. Employers and schools should not convert that limited signal into an automatic accusation.¹

Light editing may leave enough of the pattern to detect, while a complete rewrite can remove it. TechCrunch notes that Anthropic had not initially specified how much editing would defeat the signal. That creates an asymmetric policy risk: careful human revision can make detection weaker even when AI assistance was allowed, while untouched text can be flagged without proving a rule was broken.²

For supported images and files, Claude will attach cryptographically signed C2PA metadata rather than alter the content itself. Metadata can be lost when a platform strips it, so provenance checks should preserve original files and record the transfer path. Text and file evidence should not be treated as interchangeable.¹

Organizations should disclose when watermark checks are used, retain the original sample, record detector confidence and allow a person to explain the workflow. A watermark can support an investigation when combined with version history, assignment rules and human review. It should not replace those facts.