Meta official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
Copula mixing pushes detection rates past 89% while adding zero latency to AI inference pipelines
On October 29, 2026, researchers at Meta’s FAIR lab and the Meta Superintelligence Labs published a paper on arXiv introducing TextSeal2, a next-generation watermarking technique for large language models. The system detected 89% of 200-token AI-generated outputs at a false-positive rate of just 0.1%, a result that meaningfully clears the bar set by its predecessors.
What TextSeal2 actually does differently #
The original TextSeal, released in August 2026, embedded detectable signals into AI text by manipulating how models select words during generation.
TextSeal2 sidesteps that trade-off using a technique called copula mixing. Rather than maintaining multiple separate secret keys, the system consolidates them into a single virtual key while simultaneously injecting fresh noise into the process. The practical result is a stronger, more consistent signal baked into the text, without forcing the model to slow down or produce blander outputs.
Crucially, TextSeal2 adds no measurable inference latency and preserves speculative decoding acceptance rates. Speculative decoding is a speed optimization where a smaller draft model generates candidate tokens that a larger model then verifies in batch.
The research was led by Tom Sander and Pierre Fernandez, and the team has made both the code and reproducibility scripts publicly available on GitHub.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The numbers, benchmarked #
Meta ran comparative evaluations across three models: Llama-3.1-8B, Qwen3.5-9B, and Qwen3.5-27B. For 100-token sequences at a false-positive rate of 10^-3, TextSeal2 outperformed the original TextSeal by 8 to 22 percentage points depending on the model.
The most striking result came under speculative decoding conditions on the Qwen3.5-9B model. TextSeal2 recorded an 89% detection rate on 200-token outputs at that 0.1% false-positive rate. TextSeal, by comparison, achieved 71% under the same conditions. A third competing method, textGrain, managed 67%.
TextSeal2 also supports public key sharing, meaning third parties can verify whether a piece of text was watermarked without needing access to the private key that generated it.
The bigger picture: why content provenance has become urgent #
Meta’s approach is notably broader than TextSeal2 alone. The company is separately developing a technique called radioactive watermarking, which reportedly achieves a 79% detection rate for text generated by models distilled from Meta’s own systems.
As of early November 2026, Meta had not announced any public deployments or product integrations tied to the TextSeal2 findings. The research is live, the code is public, and the rest of the industry now has a new benchmark to either adopt or beat.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.