AI Reconstructed Dead Pilots’ Voices From a Crash Investigation File
The National Transportation Safety Board was forced to temporarily shut down public access to its docket system after people used artificial intelligence to rebuild the voices of pilots who died in a plane crash, a development that raises urgent questions about what should and should not be publicly available in accident investigations.
The NTSB discovered that voices of pilots killed in a UPS plane crash had been re-created using AI and were already circulating on the internet. What made it possible was not a leaked audio file, but something far less obvious: a spectrogram.
Federal law prohibits the NTSB from including cockpit audio recordings in its docket system, but the accident docket for this flight did include a spectrogram file of the voice recorder. A spectrogram uses a mathematical process to convert sound signals into a visual image. That image, it turned out, contained more than enough data to work with.
Scott Manley, a popular YouTuber known for content mixing physics, astronomy, and video games, flagged on X that it could be possible to reconstruct audio from the megabytes of data encoded in such an image. People took note and acted on it.
Using the spectrogram alongside the publicly available transcript, individuals used AI tools including Codex to create approximations of the cockpit voice recorder audio from UPS Flight 2976, which crashed in Louisville, Kentucky, according to the NTSB. The fact that AI reconstructed cockpit voice recordings with such relative ease has put a spotlight on gaps in how sensitive investigative materials are handled.
The NTSB restored public access to its docket system on Friday but kept 42 investigations closed pending review, including the one related to Flight 2976.
The incident is a stark reminder that in the age of AI, even a visual file is never just a visual file.
