This paper argues that audio deepfake detectors may exploit speaker-identity cues correlated with genuine or synthetic labels, rather than relying only on synthesis artifacts. It introduces the Identity Sensitivity Score (ISS), an inference-time, label-free diagnostic that measures how detector outputs change across reference speaker contexts. Across two detectors and two datasets, misclassified utterances had ISS values 29 to 52 times higher than correctly classified ones, with ISS reaching up to 0.954 AUC for predicting errors. In a voice-conversion test involving 500 utterances, ISS-flagged samples showed 19 to 30 times larger detector-score shifts.
No heat snapshots are available in the last 24 hours.