Witness Sensing for Verifying the Human Origin of Digital Media

As generative AI systems become increasingly capable of producing synthetic speech, images, and video that are indistinguishable from human-generated content, existing media authentication tools, largely based on post-hoc AI detection, are struggling to keep pace. In this work, we introduce witness sensing, a proactive framework for ver- ifying the human origin of digital media. Witness sensing leverages auxiliary sensor streams that are physically and temporally coupled to the act of content creation, providing a causal link to the generative process. These co-captured signals serve as hardware-based evidence of provenance. We implement this concept in the context of verifying the human origin of speech, using a radar sensor to capture vocal fold–driven neck vibrations alongside a microphone as a person is speaking. We develop a cross-modal align- ment model that maps radar signals to a filtered represen- tation of the speech waveform, and validate our approach on a dataset of 66 participants. Our analysis shows that, after isolating the vocal-fold excitation from speech via a physics-based model, the resulting signal aligns far better with the radar than raw audio stream. This is evidence that radar is capturing the same biomechanical source and thus serves as a viable physics-grounded witness sensor.