I Hear, Therefore I Trust: A Socio-Technical Investigation of Humans as Synthetic Speech Detectors

Researchers conducted a study with 47 participants to evaluate how humans detect synthetic speech, testing detection accuracy across authentic, fully synthetic, and partially synthetic utterances under various trust manipulation conditions. The findings reveal that humans perform poorly at detecting fully synthetic speech (below-chance levels) and that trust cues like instructional framing and provenance labeling do not significantly improve detection, though they influence detection behavior.

This socio-technical investigation addresses a critical gap in deepfake research by examining how humans actually perceive and interact with synthetic speech in realistic contexts rather than treating detection as a purely technical problem. The study's key finding—that participants detected fully synthetic speech at below-chance accuracy levels—suggests humans lack reliable intuitive mechanisms for identifying high-quality voice deepfakes, a concerning implication for security and misinformation resistance. The disconnect between implicit discrimination (visible in quality ratings) and explicit detection (failed localization) indicates participants could sense something was off but could not articulate or pinpoint the specific synthetic elements. Trust cues produced no main effects on accuracy, challenging the assumption that better information or context significantly improves human detection capabilities. This work positions synthetic speech detection as fundamentally different from other perceptual tasks, requiring more sophisticated cognitive frameworks. The research has substantial implications for media literacy, authentication protocols, and the design of human-AI collaborative systems. As synthetic speech technology becomes increasingly sophisticated, relying on human perception alone for detection appears insufficient, pointing toward hybrid approaches combining automated detection with user awareness training. The study's methodology—combining behavioral measurement with subjective quality ratings—provides a model for understanding human factors in adversarial AI environments. Moving forward, developers and platforms must acknowledge that human detection capabilities have inherent limitations and design accordingly, whether through technical safeguards, verification mechanisms, or honest transparency about detection challenges.

• →Humans detect fully synthetic speech at below-chance accuracy levels, indicating high-quality voice deepfakes can fool human perception
• →Trust manipulation cues like instructional framing and provenance labeling influence detection behavior but do not improve accuracy
• →Implicit discrimination in quality ratings suggests humans detect synthetic speech subconsciously even when explicit detection fails
• →Utterance class (authentic vs. synthetic) is the primary determinant of detection accuracy, not contextual trust factors
• →Human-only reliance for synthetic speech detection is insufficient; hybrid systems combining automation with human oversight are necessary