The NTSB Published a Picture. People Turned It Back Into Audio of Dead Pilots.
The NTSB publishes pictures of what cockpit voice recorders recorded. AI can now turn those pictures back into sound.
On May 19 and 20, the National Transportation Safety Board held an investigative hearing into the crash of UPS Flight 2976 — a Boeing MD-11F that came apart shortly after takeoff from Louisville Muhammad Ali International Airport on November 4, 2025, killing three crew members and eleven people on the ground. As part of the docket it releases during major investigations, the agency published a PDF containing a spectrogram: a visual representation of cockpit audio, generated from the cockpit voice recorder during the final 30 seconds of flight. On May 22, NTSB took its entire public docket system offline. According to Ars Technica, users on social media had already reconstructed what the audio sounded like — one person said it took about ten minutes using OpenAI Codex, a code-writing AI tool, to produce a rough approximation of the transcript. CNN reported the reconstruction covered the last 30 seconds of the flight.
The agency acknowledged the problem without knowing how it had happened. An NTSB spokesperson told CNN nobody was aware that you can recreate audio from a picture. NTSB Chairwoman Jennifer Homendy called the practice deeply troubling. The docket system came back online on May 22, but 42 investigations remained closed, including UPS Flight 2976, pending review. It is not yet clear whether the spectrogram was released as part of routine docket procedure and discovered later, or whether the agency identified the reconstruction risk during its own review before the public noticed — a distinction that affects how the agency characterizes its own failure.
Federal law has prohibited the public release of cockpit voice recorder audio since 1990, when Congress enacted 49 U.S.C. § 1114 specifically to protect the privacy of flight crews. The law assumed that if the audio itself never left NTSB's hands, the conversation stayed private. It did not account for a visual representation of that same audio being published as routine investigative documentation — or for a 1984 algorithm that makes it reversible.
The reconstruction technique is called Griffin-Lim synthesis, named for Daniel Griffin and Jae Lim's 1984 paper on estimating phase from an amplitude spectrogram. A spectrogram encodes sound as color and brightness across time: louder frequencies appear brighter, quieter ones darker. Griffin and Lim showed that you can recover something close to the original audio from that visual representation alone — even without perfect phase information — by iteratively guessing and refining the phase signal until the resulting audio, when converted back to a spectrogram, matches the image. Open-source implementations have existed on GitHub for years. What changed is that a code-writing AI can assemble the pipeline in minutes, from image to audio, without requiring the user to understand the underlying signal processing.
The Louisville Courier-Journal, reporting on the NTSB's own announcement, said the recreations were made using "image recognition and computational methods" applied to the visual representation of sound waves released as part of ongoing investigations. The NTSB release confirmed that it does not release cockpit audio recordings and described the reconstructed audio as "approximations." That description — "image recognition and computational methods" — is precisely what Griffin-Lim synthesis and its equivalents are: a pipeline from visual representation to reconstructed audio. Other investigative agencies and court systems use similar visual documentation of audio evidence, meaning the reconstruction exposure is not unique to the NTSB docket.
The format assumption embedded in the 1990 law is now obsolete in a specific, documentable way: a privacy protection built around the physical format of a recording — audio stays private because audio isn't released — does not survive the existence of a lossless visual representation of the same information. Spectrograms are standard documentation in NTSB investigative dockets. They are easier to include in a public filing than raw audio, and until this week they were not treated as sensitive.
Whether the answer to this vulnerability is a technical block — stripping spectrograms from public dockets, redacting the frequency bands that carry voice content — or a legal fix — extending the 1990 privacy statute to cover any processed derivative from which audio can be approximately reconstructed — both options require identifying the full scope of what is already in public filings and what other agencies face the same exposure. Records-management experts and transparency advocates have begun raising the question: if an agency can satisfy a disclosure requirement by releasing a visual representation of a protected recording, rather than the recording itself, has it actually complied with the law's intent, or found a convenient workaround?
The 1990 Congress did not have to answer that question. Today, it is unavoidable. The law protected audio because audio was the artifact. The algorithm does not distinguish between artifact and representation — it just works on whatever is in front of it.