New AI Watermark Detects 89% of Model Text
Meta's TextSeal2 avoids the output trade-off of its predecessor, but it remains a research method with no announced public deployment.
Updated on Nov. 2, 2026 in Artificial Intelligence
Live Poll
Do you trust that AI-written text can be reliably identified as machine-generated?
Meta's FAIR lab and Meta Superintelligence Labs published a paper on arXiv on October 29, 2026, introducing TextSeal2, a watermarking technique for large language models, and released its code and reproducibility scripts publicly on GitHub. The work is research-stage: Meta has not announced public deployments of TextSeal2.
Why it matters
Watermarking aims to let third parties tell whether text was produced by an AI model, and the original TextSeal relied on manipulating word selection, which imposed a trade-off in output quality. TextSeal2 is positioned as a way to strengthen the embedded signal without that trade-off, a direction that matters as detection of machine-generated text becomes a standing question for model developers and the platforms that host their output.
The evaluation is a benchmark comparison on three open models, measuring detection rate at fixed false-positive rates. TextSeal2 reached 89% detection on 200-token outputs at 0.1% false positives, against 71% for the original TextSeal and 67% for textGrain, but whether these figures hold on other models or real-world text is not established.
The players
Meta FAIR
Meta's fundamental AI research lab, known for open model releases and research into large language model training and evaluation.
Meta Superintelligence Labs
Meta's advanced AI research group, which co-published the TextSeal2 paper alongside FAIR.
The details
TextSeal2 consolidates multiple separate secret keys into a single virtual key, using a technique the researchers call copula mixing, a statistical method for combining dependent variables. It also injects fresh noise into the generation process to strengthen the embedded signal, which the researchers say adds no measurable inference latency and preserves speculative decoding acceptance rates, meaning the speed-up technique used to generate text is not degraded. The system also supports public key sharing, so third parties can verify a watermark without holding the private key.
Timeline
August 2026: the original TextSeal watermarking technique was released.
October 29, 2026: Meta researchers published the TextSeal2 paper on arXiv.
Early November 2026: Meta had not announced public deployments of TextSeal2.
The Tech Race
TextSeal2 is a follow-on to Meta's own August 2026 TextSeal release, placing the work within an internal research line that is being refined in public. The paper reports gains over the original TextSeal and over textGrain, a competing watermarking method, on the same benchmark conditions.
TextSeal2 is a research method and is not yet available in any Meta product, so there is no current change for users of Meta's services. Its code and reproducibility scripts are public on GitHub, which means developers and researchers can test it now, while any timeline for use in deployed systems has not been announced.
The takeaway
Watch whether Meta announces a public deployment of TextSeal2, since the paper stops short of one. The released GitHub scripts let outside researchers check the 89% detection figure on other models, which would test whether the gains generalize.
Further reading
For more on how model developers are approaching output tracing, see the Artificial Intelligence section.
Source note: This article includes information reported by Crypto Briefing.
Live Poll
Do you trust that AI-written text can be reliably identified as machine-generated?
