The Law Said No. AI Said Watch This.
On November 4, 2025, UPS flight 2976 — an MD-11F cargo aircraft — suffered a structural failure during takeoff from Louisville, Kentucky. The engine physically separated from the wing. Three crew members died. Twelve people on the ground were killed. Twenty-three more were injured.
Federal law, written in 1990 after pilots fought back against the public broadcasting of cockpit audio from the 1988 Delta Flight 1141 crash, prohibits the NTSB from releasing any cockpit voice or video recordings. The law exists because pilots agreed to be recorded at work on the condition that their final moments wouldn’t be broadcast for entertainment. It’s been the foundation of crash investigation cooperation for 35 years.
On May 19 and 20, 2026, the NTSB held a two-day investigative hearing for UPS 2976. They released thousands of pages of reports, a video of the engine separation, a written transcript of cockpit audio — all standard practice. They also released a PDF containing a spectrogram: a visual representation of the last 30 seconds of cockpit audio.
Spectrograms have been part of NTSB dockets for years. They’re just pictures. Pixels. You can look at them and see frequency patterns, but you can’t hear anything. That was the whole point — transparency without the audio.
Then someone on X spent 10 minutes with OpenAI’s Codex and turned the picture back into sound.
🔍 THE BOTTOM LINE
A federal law protecting pilots’ final words survived for 35 years because spectrograms were “just images.” AI made that distinction collapse in 10 minutes. Every government database containing signal data just became a potential audio recorder.
How It Happened: Griffin-Lim Meets GPT
The underlying technique isn’t new. The Griffin-Lim algorithm, published in 1984 by Daniel Griffin and Jae Lim, reconstructs audio signals from spectrograms through iterative phase estimation. Python implementations have been on GitHub for years. The math has always been there.
What changed is accessibility. Previously, reconstructing audio from a spectrogram required signal processing expertise, custom code, and patience. Now, an AI coding agent can write the code, run it, and hand you a .wav file in the time it takes to make coffee.
One X user reported taking 10 minutes with OpenAI’s Codex to reconstruct rough audio from the NTSB spectrogram. The reconstructed recordings — the final 30 seconds of three pilots struggling with a failing aircraft — spread across X and Reddit.
The spectrogram itself also circulated widely on social media, meaning the raw material for further recreations is permanently out there.
The NTSB’s Response: Pull Everything
On May 21, the NTSB announced its entire docket system — the public database containing accident reports for all civil transportation investigations — was “temporarily unavailable.”
“The NTSB is aware that advances in image recognition and computational methods have enabled individuals to reconstruct approximations of cockpit voice recorder audio from sound spectrum imagery released as part of NTSB investigations, including the ongoing investigation of the crash last year of UPS flight 2976 in Louisville, Kentucky. The NTSB docket system is temporarily unavailable as we examine the scope of the issue and evaluate solutions.”
That’s not just the UPS 2976 docket. That’s every investigation. Car crashes. Train derailments. Pipeline incidents. All offline while the agency figures out what to do.
Ben Berman, a former NTSB accident investigator and United Airlines 737 pilot, told Ars Technica: “It’s been an important factor for decades in having airline pilots be willing to have their voices recorded at their normal workplace, day in and day out, with the threat of being killed during their workday. People are horrified with the idea of their last moments being made public and used for anything other than accident investigation.”
Berman also said he was shocked it was even possible: “All kinds of things are possible now.”
The Bigger Problem: It’s Not Just Plane Crashes
Here’s what makes this story uncomfortable: the NTSB was doing nothing wrong. Spectrograms were a reasonable compromise between transparency and privacy. The data was always public. The law specifically banned audio, not visual representations of audio. Nobody at the NTSB was careless — they were following decades of established practice.
AI just changed the rules without telling anyone.
And if spectrograms can be turned back into audio, what else? Medical imaging data. Seismic recordings. Any scientific dataset where visual representations contain enough information for a model to reconstruct the underlying signal. The pattern is the same: government agencies publish derived representations of sensitive data because the raw data is protected. AI erodes the gap between representation and source.
This isn’t an NTSB problem. It’s an every-agency-with-sensitive-data problem.
What Happens Next
The NTSB has a few options, and none are great:
- Stop publishing spectrograms. Simple, but reduces transparency. Investigators and researchers use these for legitimate analysis.
- Downsample or obfuscate spectrograms. Remove enough resolution that reconstruction degrades. But this is an arms race — AI models keep getting better at filling in missing data.
- Encrypt or restrict access. Defeats the purpose of a public docket system.
- Accept that the 1990 law is now unenforceable. The most honest option, and the most uncomfortable one.
The NTSB declined to provide additional comment beyond its statement, saying it would share updates through its website and X account. The docket system remains offline.
What This Means for NZ
New Zealand’s Transport Accident Investigation Commission (TAIC) follows similar principles — it publishes investigation reports but protects cockpit audio and other sensitive recordings. TAIC will face the identical problem the moment someone runs a spectrogram from a NZ crash investigation through an AI model. If the NTSB — with vastly more resources — can’t solve this, TAIC certainly can’t.
More broadly, NZ’s AI strategy focuses on accelerating adoption. It doesn’t address what happens when AI capabilities outpace the assumptions built into privacy law. The NTSB situation is a case study in exactly that gap.
❓ Frequently Asked Questions
Q: Did the NTSB break the law by publishing the spectrogram? No. Federal law (49 USC §1114) specifically prohibits releasing cockpit voice or video recordings. A spectrogram is a visual representation, not a recording. The law didn’t anticipate that the distinction would become meaningless.
Q: How accurate are the AI recreations? They’re described as “approximations” and “rough audio.” The Griffin-Lim algorithm produces recognizable speech but with artifacts. The emotional impact of hearing a pilot’s reconstructed voice in their final moments doesn’t require audiophile quality.
Q: Can the NTSB stop this from happening again? It can stop publishing spectrograms, but it can’t un-publish the ones already circulated. The UPS 2976 spectrogram is permanently on social media. Any future solution has to assume the data is already out there.
Q: Is this an OpenAI/Codex problem? No. The technique (Griffin-Lim) predates modern AI by decades. Codex made it trivially accessible, but any capable coding model could do the same. Banning one model doesn’t fix the underlying issue.
🔍 THE BOTTOM LINE
A 1990 law protecting pilots’ final words survived for 35 years because spectrograms were “just images.” AI turned images back into audio in 10 minutes. The law didn’t change. The technology did. And every government agency that publishes derived representations of protected data — medical, defence, intelligence — just got the same wake-up call the NTSB got.
SOURCES
- Engadget — People used AI to recreate the voices of pilots killed in a plane crash
- Ars Technica — AI users re-create dead pilots’ voices from crash investigation docs
- CNN — Plane crash audio recreated from investigation documents
- NTSB — Docket system announcement
- 49 USC §1114 — Federal law on cockpit voice recorder disclosure
