The NTSB tries to keep cockpit audio recordings private. AI is making that harder

The National Transportation Safety Board (NTSB) faced an unexpected challenge when artificial intelligence tools allowed the reconstruction of cockpit audio recordings from publicly available spectrograms. The incident began with investigators examining a mysterious high-pitch ringing sound in the cockpit voice recording of UPS Flight 2976, which crashed shortly after takeoff in Louisville, Kentucky, in May 2026, killing 15 people, including all three pilots. NTSB investigators had published spectrograms—visual representations of audio frequencies—as part of the investigation. Engineer Scott Manley noticed these images and suspected they could be used to reconstruct the original audio. Within days, individuals like John McElhone, who runs a small electrical turbine company, used AI tools to convert the spectrograms back into audio in under ten minutes. Though McElhone did not share his findings, others released reconstructed audio online, exposing sensitive cockpit recordings. The NTSB responded by temporarily removing public access to all crash investigation dockets while reviewing materials and evaluating new safeguards. The agency, legally barred from releasing cockpit recordings, stated it had long-standing procedures to protect such audio but needed to adapt to AI risks. While most dockets were later restored, 41 remain under review to prevent further unauthorized reconstructions. The incident highlights how AI advancements can bypass traditional privacy measures, even for agencies bound by strict confidentiality laws. The NTSB’s swift action underscores the need for updated protocols to safeguard investigative data in an era of rapidly evolving digital tools.