Researchers behind TextSeal2 built a hidden marker for chatbot writing that spots 8 to 22 percentage points more short AI replies without slowing them down or making them sound more repetitive.
Key Takeaways
TextSeal2 detects 8 to 22 percentage points more short model outputs than prior systems while adding no measurable inference latency
The paper said TextSeal2 keeps its signal intact under speculative decoding, a speedup method many production systems use
TextSeal2 merges multiple keys into one virtual key through a copula-based method to keep detection strong without making text feel more repetitive
Rejected draft tokens are replaced so the watermark key is preserved, keeping acceptance rates unchanged and not increasing latency
The paper said TextSeal2 keeps its signal intact under speculative decoding, a speedup method many production systems use to draft and verify tokens faster. An AI watermark is a hidden statistical pattern in generated text.
It is not a visible label. It changes token selection in a way a detector can test later, if it holds the right key.
Multi-turn safety and provenance have both become harder as serving stacks grow more complex.
The same week brought fresh evidence that LLM safeguards can weaken once outputs pass through longer chains of prompts and tools.
AI Watermark Meets A Real Deployment Problem
Many watermark ideas work in laboratory settings, then weaken when model serving stacks change. One problem is speculative decoding, where one model drafts tokens and another verifies them to speed up output.
That process can scramble or dilute the hidden pattern older watermark methods rely on.
If the verification stage changes enough tokens, the detector later struggles to tell machine text from human text.
TextSeal2 tries to fix that exact failure mode. The paper says rejected draft tokens are replaced in a way that preserves the watermark key, so acceptance rates stay unchanged and latency does not rise.
From TextSeal To TextSeal2
The new method builds on the earlier TextSeal system, which already tried to preserve output diversity better than blunt watermarking schemes.
The weakness, the authors argue, was that the detector averaged across separate keys without knowing which key was active for each token.
Also Read: Free AI From A Startup Reads Photos And Text On Ordinary Laptops
TextSeal2 merges multiple keys into one virtual key through a copula-based method. That statistical step is dense on paper, but the practical aim is simple.
It keeps detection strong without making the text feel more repetitive.
That matters for labs and model hosts. A watermark nobody can detect is useless, but a watermark that may make responses noticeably worse is also a nonstarter.
The Oct. 29 paper arrives after a year in which AI labs shifted from watermark theory toward operational questions such as latency, serving compatibility and adversarial rewriting.
That is why speculative decoding matters here more than abstract benchmark gains.
How The Detector Stays Hidden
TextSeal2 uses randomness to steer token choices among many plausible next words, rather than stamping obvious markers into output. The detector later tests whether the resulting token pattern is too consistent with the secret key to be human chance.
Because the model still picks from valid tokens, the text can remain natural.
The challenge is balancing three goals at once, keeping fluency, keeping speed and keeping enough signal for later detection.
For model providers, the biggest significance is operational. Open systems increasingly need a way to distinguish generated text after the fact, especially when outputs circulate without metadata or when users copy content into third-party tools.
The next step is external adoption, not arXiv gains.
Labs will need to show whether an AI watermark like TextSeal2 survives adversarial rewriting, multilingual output and long production chains that include summarizers, editors and retrieval steps.
Read Next: OpenAI Boss Sam Altman Says AI On Your PC Wastes Precious Memory