Proprietary · Forensics · EN

Watermark Forensics™

LLM Statistical Watermark Detection & Neutralization

Watermark Forensics™ detects and neutralizes invisible statistical watermarks that large language models (GPT-4, Claude, Gemini) embed in their output — specifically token-selection biases, green-list/red-list token distributions, and n-gram frequency anomalies.

How It Works

Watermark Signature Detection

Scans token-frequency distributions for the characteristic bimodal pattern of green-list watermarking (Kirchenbauer 2023 scheme and variants). Also detects self-hash and distortion-free watermarks.

Token Bias Analysis

Maps per-token selection probabilities against an estimated baseline LM distribution. Tokens with statistically anomalous selection rates (p < 0.01 vs baseline) are flagged as watermark carriers.

Targeted Neutralization

Replaces flagged tokens with semantically equivalent alternatives drawn from the opposite (red-list) token pool — eliminating the statistical signal while preserving meaning.

Forensic Report

Generates a detailed forensic report: detected watermark type, strength score (0–1), estimated source model, flagged token positions, and confidence interval.

Key Metrics

3 (KGW, SelfHash, DF)

Watermark schemes detected

91% (KGW scheme)

Detection accuracy

87%

Neutralization rate

< 4%

False positive rate

200 tokens

Min text length

Clean text + forensic report

Output

Use Cases

Detecting whether a document was generated by a watermarked LLM
Neutralizing watermarks before publishing AI-assisted content
Forensic analysis of suspected AI-generated documents
Research into LLM traceability and detection evasion

Try Watermark Forensics™ Now

Open Watermark Tool API Docs All Technologies Pricing