You can detect fake audio spliced into real recordings without training a detector

Researchers found that speech AI models already contain hidden forensic signals: genuine speech flows smoothly through the model's internal representations, while splice points create abrupt jumps. A new detection method reads these jumps directly from existing AI models without any training, labeled data, or model modification—and catches most partial deepfakes as well as supervised detectors that required extensive labeled examples.

For years, detecting deepfakes required building a new detector every time someone released a new synthesis tool, which meant the arms race always favored the fakery side. This removes that requirement—the detector works on future synthesis methods it has never seen because it reads a forensic signal that's baked into how speech models work. What matters in practice: you stop needing to retrain detectors constantly, and the detection gets harder to evade by simply training on different synthesis pipelines.