The NTSB publishes accident investigation dockets online - thousands of pages of evidence, transcripts, and technical data that researchers, journalists, and families use to understand what went wrong. Those dockets sometimes include spectrograms: visual charts that map sound frequencies over time, released as images rather than audio files. The agency has treated this as a reasonable middle ground for years. You can see the data without hearing the voices of the people who died.
That assumption broke in May 2026. Researchers used AI to reconstruct audio from spectrogram images published in the NTSB's public docket system, effectively recreating the voices of pilots killed in crashes. The agency responded by temporarily taking its entire docket system offline.
How a Picture Becomes a Voice
A spectrogram isn't just a visualization - it encodes most of the information needed to reconstruct the original audio: frequencies, amplitudes, and timing all mapped across a two-dimensional image. Traditional methods of reversing a spectrogram back into sound produced degraded, often unintelligible results. AI-based vocoder models (software that synthesizes speech from spectral data) have changed that. The reconstruction is now intelligible enough that the NTSB felt it had a real problem.
The technique itself isn't new. Producers have used AI spectral reconstruction to restore old recordings for years. Applying it to aviation accident data hits a specific legal and ethical wall: cockpit voice recorder audio is federally restricted from public release specifically to protect the privacy of deceased crew members and their families. The NTSB thought publishing spectrograms instead of the audio files sidestepped that restriction. It didn't.
What Comes After the Docket Goes Dark
Taking the docket system offline is understandable as an immediate response, but it's expensive. Independent aviation safety researchers, accident investigators, and journalists worldwide rely on those records. The NTSB's public docket is one of the most transparent accident investigation systems in the world - pausing it delays legitimate safety work.
The harder question is what changes permanently. Scrubbing spectrograms from future publications is straightforward. Auditing years of historical dockets is not. And the underlying problem is broader than spectrograms: any detailed data representation that was considered "safe" to publish under old assumptions can become unsafe when AI lowers the reconstruction threshold. That includes other waveform visualizations, redacted documents where physical indentations from writing remain legible, and metadata embedded in exported files.
This is a specific and unusually dramatic version of a problem security researchers have discussed in the abstract for years. The NTSB is not the last institution to discover that their privacy model assumed capabilities that no longer reflect reality.