← Accueil actu | Salle du marché

Deepfake Detectors Lose 49% Accuracy Against Latest AI Video Generators

2026-09-14 · Source:Unite AI

Deepfake Detectors Lose 49% Accuracy Against Latest AI Video Generators

A Ukraine-led study finds legacy deepfake detectors failing against generators like Alibaba's Wan 2.2, Hunyuan Video, Grok Imagine and Kling, with accuracy down 49% in a year.

A Ukraine-led research paper has found that the older generation of deepfake video detectors performs very poorly against current state-of-the-art generative AI video. The headline measurement is stark: detection accuracy has fallen 49% across a single year of progress in AI video generation. Crucially, the weakness holds regardless of whether the synthetic footage was produced by open-source models running locally or by closed, API-only commercial services.

On the open side, the study cites locally installed generators including Alibaba's Wan 2.2 and Hunyuan Video. On the closed side, it names API-only models such as Grok Imagine and Kling. That breadth matters for anyone operating media-verification tooling: the detection failures are not confined to one vendor, one architecture or one distribution channel, but span downloadable open weights and hosted services alike, meaning no procurement choice among today's generators buys immunity from fooling legacy detectors.

The evidence base is a benchmark the authors call DF26. To illustrate the challenge, the paper shows three samples drawn from its 'Direct to camera casual' category — synthetic clips in which an on-camera subject addresses the viewer in an informal, everyday register. That style is precisely the format hardest to police with rules or metadata checks, and the samples underscore how far current generators have moved from the uncanny artifacts older detectors were trained to catch.

The paper is not the first warning of its kind. Back in 2024, another paper had already pointed in the same direction, and the intervening twelve months of generator progress appear to have widened the gap rather than closed it — a 49% drop in detector accuracy. For enterprises and platforms relying on automated video verification, the study implies that legacy detection stacks benchmarked against older generator families can no longer be assumed to hold against the current wave of open and closed models.

Mots-clés: Alibaba · Wan 2.2 · Hunyuan Video · Grok Imagine · Kling · Deepfake Detection


À lire aussi


萬安算交所 AIXX · AIXX · Bourse compute vérifiée · Accéder à la salle