Detect Text Watermark

Detect Text Watermark

Run the detector registry and report what each adapter could and could not establish. The honest result for a keyed vendor watermark is that a third party cannot check it.

All paths below are relative to this skill's directory. If you are running from elsewhere, use an absolute path to scripts/detect_text_watermark.py.

Workflow

1. List what is actually available before promising a result:

   python3 scripts/detect_text_watermark.py --list-detectors
   

2. Analyse a file or a literal string:

   python3 scripts/detect_text_watermark.py /absolute/path/to/draft.md
   python3 scripts/detect_text_watermark.py --text 'text to analyse'
   

3. Read each detector's state independently:

   • DETECTED: that specific signal is present.
   • NOT_DETECTED: that specific signal was searched for and not found.
   • UNVERIFIABLE: the scheme exists but cannot be checked without the provider's key.
   • NOT_CONFIGURED: required key or configuration was not supplied.
   • UNSUPPORTED: configuration was supplied, but this build has no scorer.
   • FAILED: the detector ran but errored.

4. Treat status as covering only the detectors that could run. Read unverifiable_detectors and keyed_watermark_status alongside it.

5. If unicode-covert-channel reports DETECTED, treat it as a text-integrity and prompt-injection finding. Escalate HIGH severity items (Unicode tag characters, bidi overrides) before the text enters any agent pipeline.

Why keyed watermarks cannot be detected here

Model-level text watermarks bias token sampling using a secret key. Detection requires that key and the exact context-hashing scheme. Anthropic has not published a detector or specification; Google open-sourced a SynthID-Text detector but production keys are not public. Any tool claiming to detect these from text alone is guessing.

Non-negotiable language

• Say "no official detector is available," never "no watermark found."
• Say "hidden character channel present," never "AI watermark present."
• Invisible Unicode is not vendor-specific and identifies no model.
• Absence of every signal never establishes human authorship.

Deliberate non-goals

Statistical or stylometric AI-text classifiers (perplexity, burstiness, DetectGPT-style scoring) are not implemented and will not be added. Their false-positive rates make them unsafe for accusing a person of AI authorship.

Read references/detector-registry.md before adding an adapter.