TextSeal: Localized Watermarking and Traceability Protection for Large Language Models

TextSeal is an advanced watermarking technology for large language models (LLMs). Its core features include: supporting multi-region localized detection, maintaining high detection confidence in human-AI hybrid documents; having a "radioactive" property that allows watermark signals to be transmitted during model distillation to prevent unauthorized use; being theoretically distortion-free, without affecting text quality or the model's output distribution. This technology aims to address cross-domain core issues in AI content traceability, providing reliable solutions for scenarios such as academic integrity, copyright protection, and misinformation governance.

As the generation capabilities of large language models improve, distinguishing between human and AI-created content has become both difficult and important, involving fields such as academic integrity, news authenticity, copyright protection, and misinformation governance. Watermarking technology faces three major challenges: invisibility (not affecting text quality), robustness (resisting operations like rewriting/translation), and localization ability (accurately identifying AI-generated paragraphs). Existing solutions are mostly based on vocabulary replacement or statistical feature modulation, which are easy to remove, affect quality, and cannot locate specific AI-generated parts.

TextSeal innovates based on the Gumbel-max sampling framework:

1. Dual-key generation mechanism: Restores the natural diversity of output text; multiple generations from the same prompt show significant differences but the watermark remains detectable;
2. Entropy-weighted scoring system: Focuses on positions with high information entropy (vocabulary where the model has multiple choices) to improve detection accuracy;
3. Multi-region localized detection: Divides the document into multiple regions, evaluates watermark confidence separately, and accurately locates AI-generated paragraphs instead of making an overall judgment.

TextSeal is seamlessly compatible with inference acceleration technologies like speculative decoding, with no additional inference overhead. Its detection performance surpasses Google SynthID-text, achieving a higher detection rate at the same false positive rate. It has strong dilution robustness, enabling high-confidence localization of AI segments in human-AI hybrid documents. Multi-language tests (English, Chinese, Spanish, French, German) show no perceptible quality degradation; humans cannot distinguish between watermarked and non-watermarked text. Theoretically, it is distortion-free, does not change the model's output distribution, and does not affect the accuracy of downstream tasks.

The "radioactive" property of TextSeal makes its watermark signal contagious, allowing it to be transmitted to new models during model distillation. Traditional watermarks are lost during distillation, but TextSeal can detect watermark traces in the output of distilled models, effectively preventing unauthorized model distillation and providing model owners with a technical means to track illegal derivative versions.

TextSeal is suitable for various scenarios:

• Model service providers: Automatically add watermarks to API outputs as a standard process;
• Enterprise users: Support custom keys; watermarks embedded with private keys can only be identified by the holder;
• Content moderation: Highlight suspected AI-generated paragraphs to help reviewers quickly locate key content. Deployment does not affect user experience or computing costs.

Current limitations:

1. It assumes that attackers cannot obtain the original model or watermark key; protection may fail in extreme cases;
2. Detection confidence decreases for extremely short texts (single sentences/few words);
3. Adversarial attacks specifically targeting watermarks may weaken detection effectiveness. Future directions: Optimize short text detection, address adversarial attacks, and enhance protection capabilities in extreme scenarios.

TextSeal represents an important advancement in LLM watermarking technology. Through innovations such as dual-key, entropy weighting, and localized detection, it achieves a balance between detection strength, robustness, and invisibility. The radioactive property opens up new possibilities for model traceability. As AI-generated content becomes more prevalent, TextSeal provides a technical foundation for building reliable infrastructure for the digital content ecosystem.