text-watermarking

[ACL 2024] Can Watermarks Survive Translation? On the Cross-lingual Consistency of Text Watermark for Large Language Models

Local, evidence-first inspection for AI-generated text and files. Detection is not interpretation.

Local-first toolkit for detecting AI-generated text and disrupting classical detectors plus statistical LLM watermarks (SynthID-Text / tournament-sampling style).

Geometry of meaning-preserving rewriting paths and the erosion of statistical text watermarks.

语义保持型中文文本 / Python 代码水印：同义词替换与标识符改写嵌入，盲提取、碎片恢复、seed 密钥。Semantic-preserving watermarking for Chinese text & Python code — blind extraction, fragment recovery, HMAC seed keying.

A research implementation of statistical text watermarking for large language models using Plug and Play Language Models (PPLM). This system enables detectable watermark embedding through direct logit perturbation during inference without modifying the base model weights.

An honest in-browser model of how AI-text watermarks (SynthID-style) are planted with a secret key and detected — showing why the wrong key, and watermark removal, hit a wall.

Browser-based watermark-vs-signature demo — real SynthID-Text keyed g-value detection, tournament sampling over pinned GPT-2 distributions, entropy limits, attack sweeps, and ECDSA P-256 over a C2PA-shaped manifest. Change only the key and the evidence vanishes.

Reproducible SynthID-Text implementation for vLLM, datasets, detectors, and evaluation results