TextSeal2 aims to preserve AI text watermarks during faster decoding

What the study found

In an arXiv preprint, Tom Sander and colleagues describe TextSeal2, which combines multiple watermark keys into a single virtual key and couples it with fresh noise. They say the approach avoids weakening detection when additional keys are merged, including publicly shared keys.

The authors evaluated the method on three language models. At equal output-token diversity, they report detecting 8 to 22 percentage points more 100-token answers than the earlier TextSeal method, at a false-positive rate of 0.1%. They report detection comparable to textGrain without post-logits optimization.

Under speculative decoding, which uses draft tokens that are then verified, TextSeal2 detected 18 and 22 percentage points more 200-token answers than TextSeal and textGrain, respectively, at the same acceptance rate. The authors say their verification method also watermarks tokens that replace rejected drafts.

Why it matters

Watermarks are intended to help identify model-generated text. The authors argue that TextSeal2 maintains its signal while preserving output diversity and the target model’s output distribution. They also report no measurable inference-latency increase, making the approach potentially relevant to efficient text generation.

Limits and unknowns

These findings come from a preprint abstract. It does not identify the three models, describe the evaluation datasets or give detailed latency measurements. It also does not report robustness to editing or deliberate watermark removal, so the abstract alone cannot establish how the method would perform in those conditions.