The NTSB temporarily blocked public access to its accident docket system after AI tools reconstructed the voices of pilots killed in a UPS crash. An unforeseen technical gap reveals how far modern AI can go when given data never designed to be a privacy risk.
Key Takeaways
- AI tools rebuilt pilots’ voices from a spectrogram published in a public accident docket
- The NTSB restored access but 42 investigations remain closed pending review
- Federal law bans cockpit audio from public dockets, but spectrograms were never covered
Have an AI Sum Up This Article
ChatGPTA Spectrogram, a Transcript, and Codex
The public accident docket for UPS Flight 2976, which crashed in Louisville, Kentucky in 2025, included a spectrogram from the cockpit voice recorder. A spectrogram is a visual representation of a sound signal, encoding frequencies, amplitudes, and time. On the surface, it looks like an image. In practice, it is a structured audio signal compressed into a visual format.
Users combined that spectrogram with the publicly available transcript from the same docket. Using Codex and other AI tools, they reconstructed audio from the megabytes of data encoded in the image. The output: a recreation of the voices of the pilots killed in the accident. For context, see our earlier piece on Horizon: DuckDuckGo Up 30%: Users Reject Google’s Forced AI Search. For context, see our earlier piece on Horizon: DuckDuckGo Up 30%: Users Reject Google’s Forced AI Search.
Scott Manley, a science and technology YouTuber, was among the first to document the method. According to him, “it could be possible to reconstruct audio from the megabytes of data encoded in that image.” The combination of a precise textual anchor and a high-resolution spectrogram provided enough signal for an AI model to perform the voice synthesis.
This reconstruction would not have been possible from the text transcript alone. It was the spectrogram that made it feasible: a file no one had identified as a potential audio extraction vector. Its presence in a supposedly harmless public document exposed a gap in a regulatory framework built decades before modern AI existed.
The incident points to a shift that public institutions are only beginning to grasp. Data treated as a secondary visual representation of a signal can become fully exploitable primary data in the hands of the right AI tools. The spectrogram was not the audio, but it contained enough information for a modern model to cross that line.
The NTSB Responds: 42 Investigations on Hold
After the discovery, the NTSB immediately suspended public access to its accident docket system. Access was restored the following Friday, but with a significant constraint: 42 investigations remain closed pending a full review. The docket for UPS Flight 2976 is among those that remain locked.
Federal law already explicitly prohibits the inclusion of cockpit audio in public dockets. That rule exists to protect victims’ families and preserve the integrity of investigation procedures. Spectrograms, as visual representations, were never covered by that prohibition. That regulatory blind spot is exactly where AI exploitation found its opening.
The NTSB did not detail its internal review process or provide a timeline for the 42 affected dockets. Each file will likely be audited for the presence of spectrograms or other visual representations that could be exploited in similar ways. No reopening procedure has been publicly communicated.
This rapid response is a first for the agency. It shows that an institution historically focused on technical aviation safety investigation is now forced to actively monitor AI use against its own digital archives. Documentary transparency, a cornerstone of the NTSB’s model for decades, is now in direct conflict with AI reconstruction capabilities.
Also on Horizon:
- Meta Lays Off 8,000 Workers to Fund AI
- OpenAI IPO Is Targeting September 2026
- Anthropic Is About to Post Its First Profitable Quarter
Regulation, Privacy, and AI: A Precedent With Lasting Consequences
In the short term, families of victims in the 42 affected cases are left in uncertainty. Dockets they might need for legal proceedings or administrative processes remain inaccessible. This is not a theoretical consequence: it is a direct impact on people already dealing with loss.
For aviation safety investigation teams, the issue is more structural. The NTSB’s public archives exist because transparency is considered a key factor in collective safety improvement. If that transparency becomes a vector for AI exploitation, the trade-off between openness and protection becomes radically more complex. Restricting data weakens the agency’s mission. Keeping it open exposes victims.
In the medium term, this incident will likely reshape standards for accident data publication well beyond aviation. Other safety agencies will need to audit their own digital archives in light of this case. A spectrogram was an invisible risk in 2025; in the months ahead, it could be a telemetry chart or a biometric data file.
Case law on using AI to reconstruct personal data from secondary visual representations does not yet exist. This case could become a reference point in debates about post-mortem privacy rights and the liability of platforms hosting tools like Codex. Public safety archives had never been designed to defend against this kind of exploitation.
Follow the story on Horizon.
