AI recreates voices of dead pilots, forcing NTSB to temporarily shut down public database

The National Transportation Safety Board temporarily shut down public access to its investigation database after discovering that people had used AI tools to recreate the voices of pilots killed in a UPS cargo plane crash. The synthetic voices of the deceased pilots from Flight 2976 were circulating on the internet, prompting the aviation safety agency to restrict access to its normally open records system.

The incident highlights the growing ease with which AI can be misused to create disturbing content, particularly involving real people who died in tragic circumstances. While deepfake technology has mostly grabbed headlines for creating fake celebrity videos or political disinformation, this case shows how it can intrude on sensitive investigations and potentially traumatize families of accident victims.

Federal law prohibits the NTSB from including actual cockpit voice recordings in its public docket system, which contains vast amounts of data on transportation accident investigations. However, the agency did include a spectrogram file from the flight recorder – a visual representation that converts sound signals into mathematical data displayed as an image.

YouTuber Scott Manley, whose popular channel covers physics and astronomy topics, pointed out on social media that it might be possible to reconstruct audio from the data contained in such spectrograms. That’s exactly what happened next.

Internet users took the publicly available spectrogram along with the cockpit transcript and fed both into AI tools like Codex to generate approximations of what the pilots’ final conversations sounded like during the fatal Louisville, Kentucky crash, according to the NTSB.

The agency restored public access to its docket system on Friday but kept 42 investigations closed while it reviews what information should remain publicly available. The Flight 2976 investigation remains among those still sealed.

This case reveals an unexpected vulnerability in how government agencies share investigation data. While spectrograms were likely considered safe to publish since they don’t contain actual audio, the combination of AI voice synthesis and publicly available transcripts created a workaround that the NTSB apparently didn’t anticipate.

The incident also raises broader questions about how AI tools are making it easier to create disturbing content involving real people, particularly those who can’t consent because they’re deceased. As AI voice cloning becomes more accessible and sophisticated, similar situations may arise across other industries and government agencies that publish data containing traces of human voices or speech patterns.