The National Transportation Safety Board temporarily yanked public access to its docket system after discovering that AI-generated recreations of dead pilots’ voices were circulating online. According to TechCrunch AI, the voices belonged to the crew of UPS Flight 2976, which crashed in Louisville, Kentucky last year. The agency restored access Friday but kept 42 investigations sealed pending review, including the UPS crash file.
Here’s what makes this story unusual: federal law already bars the NTSB from publishing cockpit audio recordings in its public docket. The agency followed that rule. But the file for Flight 2976 contained a spectrogram, a visual representation of the voice recorder data. A spectrogram converts sound frequencies into an image using a mathematical process. On paper, it’s a chart. In practice, it’s audio in a different costume.
How the reconstruction happened
YouTuber Scott Manley flagged on X that the megabytes of data encoded in that spectrogram image could be turned back into sound. People online took him up on it. Per TechCrunch AI, they combined the spectrogram with the publicly available crash transcript and used AI tools, including OpenAI’s Codex, to reconstruct approximations of what the pilots said in their final moments.
The results spread across the internet before the NTSB pulled the plug.
Why this matters for the AI industry
This is a textbook case of capability outrunning policy. The NTSB’s disclosure rules were written for a world where a spectrogram was just a technical artifact, readable only by acoustic specialists with niche software. AI flattened that gap. Now anyone with a chatbot and an afternoon can pull voice data out of a chart that was never meant to be heard.
A few implications worth flagging:
- Data minimization rules need rewriting. Agencies that publish derivative files (spectrograms, transcripts, telemetry plots) now have to assume those files can be reverse-engineered into the underlying source material. The line between “public metadata” and “protected raw recording” is gone.
- Voice cloning has crossed into forensic territory. Until recently, AI voice resurrection was a novelty (think de-aged actors, Anthony Bourdain narration). Using it on crash victims’ final words is something else entirely. Expect families, unions, and regulators to push back hard.
- OpenAI’s Codex got named. When a tool gets cited by name in an NTSB statement about misuse, that’s the kind of attention vendors don’t want. Watch for tightened usage policies around audio reconstruction prompts.
What comes next
The NTSB review of those 42 sealed investigations will probably set a new disclosure standard, and not just for aviation. Any federal agency that publishes signal data, medical imaging, sensor logs, or any file that can be inverted into something more sensitive, is now on notice. The legal definition of “audio recording” may need to expand to cover anything an AI can turn into one.
The broader lesson is one this industry keeps relearning: when you ship a public dataset, you’re not just shipping what’s in it. You’re shipping everything a sufficiently capable model can extract from it. That math changes every six months.
Full details at the original TechCrunch AI report.