A browser plugin now lets journalists and fact-checkers detect AI-generated audio without uploading a single byte to an external server.
Researchers published a lightweight deepfake detection model that runs entirely in the browser. The approach pairs a truncated self-supervised learning backbone with a logistic classifier — a combination that beats the standard AASIST benchmark by 10% on accuracy while cutting inference time by 40%. The plugin code is open-source and available on GitHub. Unlike commercial tools, it requires no cloud connection.
For journalists and fact-checkers, the cloud dependency in most deepfake tools is a real problem: sending an unverified audio clip to a third-party server can expose a source before a story is even published. A local model sidesteps that entirely. That matters more now that audio deepfakes have moved from a novelty to a routine disinformation tactic.
The research is a reminder that faster and smaller do not have to mean worse — the truncated backbone here beats heavier cloud-dependent systems on the accuracy metrics that matter most for this use case.