As large language models (LLMs) like ChatGPT become more common, detecting AI-generated text reliably and without altering the quality of the output is a growing challenge. Researchers have now developed TextSeal2, a new watermarking technique designed to embed hidden signals in AI-generated text without distorting it or slowing down the generation process. This breakthrough could help distinguish human-written content from AI-generated writing more accurately, which is important for issues like misinformation, content authenticity, and copyright enforcement.
Key Takeaways
- TextSeal2 embeds watermarks in AI-generated text without affecting how natural or diverse the output sounds.
- It maintains its watermark signal even when the AI uses speculative decoding, a technique that improves text quality by considering multiple possible token choices.
- The method merges multiple watermark keys into one “virtual” key using a statistical tool called a Gaussian copula, improving detection accuracy.
- Compared to previous approaches, TextSeal2 detects AI-generated text more reliably—up to 22% better on longer passages—without adding processing delays.
Watermarking in AI-generated text is like embedding a secret signature that can be detected later to verify whether a piece of writing was produced by a machine. Earlier methods often came with trade-offs: they might make the text less natural, reduce variety in word choices, or slow down the text generation. Speculative decoding is a newer technique that lets AI models “think ahead” by generating multiple draft options and picking the best, which improves fluency but complicates watermarking.
TextSeal2 builds on a previous method called TextSeal, which used a simple “coin flip” to decide between two watermark keys for each word. However, this approach diluted the watermark signal because detectors had to average across keys without knowing which one was used, reducing detection strength. To overcome this, the researchers introduced a way to combine two or more watermark keys into a single “virtual” key using Gaussian copula—a statistical method that links multiple random variables to preserve their dependence structure.
This merging means the watermark detector no longer struggles with uncertainty about which key was applied, leading to stronger detection signals. Additionally, TextSeal2 incorporates a “memoryless race” mechanism to verify watermarked drafts even when speculative decoding generates multiple token candidates. This ensures that the final output matches the original model’s behavior exactly, preserving output quality and diversity.
The team tested TextSeal2 on three different large language models. They found that at the same level of output diversity, TextSeal2 detected AI-generated text passages between 8% and 22% more effectively than the original TextSeal method. When using speculative decoding, it outperformed both TextSeal and another watermarking approach called textGrain by 18% to 22% on longer 200-token passages. Importantly, these gains came without adding any noticeable delay in text generation or requiring complex post-processing steps.
Looking ahead, TextSeal2’s ability to watermark AI-generated text without compromising quality or speed makes it a promising tool for platforms and organizations aiming to authenticate content or combat AI-generated misinformation. The approach’s compatibility with multiple watermark keys and its robustness under advanced decoding methods suggest it could adapt well as language models continue to evolve. Future research may explore integrating TextSeal2 directly into commercial AI systems or combining it with other verification techniques to enhance trust and transparency in AI-generated content.
Based on research published on arXiv by Tom Sander, Pierre Fernandez, Itai Gat et al..