Voices of Dead Pilots Resurface Online Thanks to AI, Authorities Forced to Darken 42 Investigations

In recent days, the NTSB, the U.S. federal agency for transportation safety, suspended public access to its investigative dossier system after discovering that some users had employed AI tools to reconstruct the cockpit audio of a cargo flight that crashed last November.

The incident at the center of the case is that of flight UPS 2976, an MD-11F that took off from Louisville, Kentucky, on November 4, 2025: an engine detached from the wing during takeoff, leading to the aircraft's crash. The three crew members and twelve people on the ground lost their lives. During the two-day investigative hearing that took place last week, the board released the complete dossier: thousands of pages of technical reports, a video of the engine separation, the transcript of the cockpit voice recorder (CVR), and a PDF containing a spectrogram of the recorded audio.

The spectrogram is a visual representation of sound, where the time and frequency axes describe the amplitude and distribution of the original audio. It is a simple image and this is exactly why it is published: under federal law, the NTSB cannot release the CVR audio due to the particularly sensitive nature of the contents recorded in the cockpit during the final moments of a flight. The implicit premise is that the two-dimensional image cannot be traced back to the original sound. However, as mentioned, the dossier also included the textual transcript.

Ten Minutes with Codex to Extract the Audio

NTSB doesn't release cockpit voice recorders from crashes, except in this case they've released an image of a spectrogram. I'm not sure that's a good idea since you can probably reconstruct a lot of audio from the megabytes of data encoded in this image.

pic.twitter.com/TpteaY0Tt4
— Scott Manley (@DJSnM)
May 20, 2026

YouTuber Scott Manley, known for producing educational content intertwining physics, astronomy, and video games, noted on X that the megabytes of data encoded in the spectrogram image were sufficient to reconstruct the original audio signal. A few days later, some users did exactly what Manley had hypothesized: by combining the spectrogram with the publicly available transcript in the dossier, they generated audio approximations of the last thirty seconds of the flight, as the crew attempted to maintain control of the aircraft. One user stated on X that they did it in about ten minutes using Codex, OpenAI's agent designed for programming tasks.

I was able to reconstruct the rough audio from the spectrogram, it's surprisingly not as noisy as expected, and you can make out the entire transcript fairly clearly.
I won't upload it for obvious reasons but you're right, took maybe 10 mins on Codex to get it
— John McElhone (@Johnmcelhone)
May 21, 2026

The generated clips began circulating online alongside a second audio file reconstructed from acoustic tests conducted by the NTSB on the aircraft. The board's reaction was clear: "We publish our work and have been doing so for years. No one was aware that audio could be reconstructed from an image," a spokesperson for the agency stated in a statement collected by CNN. The agency added that it is verifying whether other dossiers contain elements that, in light of these digital reconstruction techniques, might compromise the privacy of the individuals involved.

Forty-Two Investigations Still Blocked

Public access to the system was restored on Friday, May 22, but forty-two investigations remain closed pending review, including the one on flight UPS 2976. The NTSB did not provide a timeline for the review nor clarified whether future dossiers would exclude spectrograms from the materials made public.

The case illustrates a problem that the spread of generalist AI tools is amplifying: analysis and reconstruction techniques that required specific skills just a few years ago are now operationally within reach for anyone with access to an AI agent. The transformation of an image into audio signal by reversing the mathematical operations that produce a spectrogram has long been known in principle. What has changed is the ease with which this can be done: where one previously needed to tackle writing custom code, now everything can be conducted almost automatically with the help of an AI agent in just minutes.

The NTSB will now have to assess whether the benefits of transparency in the current format, designed for safety researchers, journalists, victim families, and legal experts, remains compatible with the federal rule prohibiting the dissemination of the original audio.