AI-generated text needs a hidden watermark to detect misuse. The watermark must preserve text quality and survive , a speed-up technique that generates many tokens at once.
Prior methods using two secret keys forced the detector to average scores, weakening the signal. TextSeal2 merges multiple keys into one virtual key using a Gaussian copula, and exploits memorylessness so speculative decoding proceeds without lowering draft acceptance.
At a low rate of false positives, TextSeal2 detected far more text than previous methods, matching the best detector without per token optimization. This removes a key trade-off, making robust watermarking simpler to deploy.