When planes crash, the federal government releases the transcripts. It does not release the audio. That distinction just stopped meaning anything.
On November 4, 2025, UPS Flight 2976 crashed in Louisville shortly after takeoff. The NTSB ran its usual investigation, held a public hearing on May 19, and posted its full docket online, including the Cockpit Voice Recorder Group's factual report. That report contained spectrograms, which are basically pictures of the audio waveform, sitting next to the written transcript. Within days, users online had fed the spectrograms into audio reconstruction tools and produced clips of the pilots' final words. The clips spread on social media. By May 21, the NTSB had taken its entire public investigation database offline.
A bit of background on why that's a big deal. Under federal law, the NTSB is not allowed to release CVR audio. The agency itself describes the recordings as having a "highly sensitive nature" and says the timing and content of transcript releases are "strictly regulated." A committee of representatives from the airline, the manufacturer, the FAA, and the pilots union reviews every transcript before it goes out. The audio stays sealed. The transcript goes public. That has been the bargain aviation safety has run on for thirty-five years, because it lets investigators record everything in the cockpit without families and the public ever hearing crew members in their final moments.
What broke that bargain isn't really frontier AI. The algorithm that turns a spectrogram back into sound was published in 1984. As Thomas Claburn pointed out at The Register, and as several engineers noted on Hacker News, what people were running was essentially a vibe-coded version of a forty-year-old technique. The NTSB's actual mistake was publishing the spectrograms in the first place. AI just made the reverse engineering easy enough that someone bored on a Wednesday could do it on a laptop.
An NTSB spokesperson told CNN: "We show our work and we've been doing this type of thing for years. Nobody was aware that you can recreate audio from a picture."
The legal landscape around voice cloning is moving in a direction the agency might not have thought about either. Last July, a federal judge in New York ruled in Lehrman v. Lovo that AI voice clones can count as misappropriation of identity under state civil rights law. The judge cited a 1988 case involving Bette Midler and Ford, writing that "the human voice is one of the most palpable ways identity is manifested." A separate lawsuit against ElevenLabs argues that the accuracy of its voice clones is itself evidence that the company trained on copyrighted recordings without permission.
That framework is being built for the living. The dead UPS pilots have no plaintiffs, no estates filing complaints, no civil rights claim to make. Their voices were reconstructed from federal records that the government published itself.
Locking the docket solves the immediate embarrassment but not the actual problem. The spectrograms have already been downloaded. The reconstructions already exist. The ElevenLabs lawsuits show what happens next, which is that even when a company pulls a voice off its platform, users rebuild it under a different name within days. What aviation safety just learned is that the line between transcript and audio, the one a federal statute spent thirty-five years defending, was holding up only because recovering the audio used to be hard. It isn't anymore. Every agency that built its disclosure rules around the assumption that raw data is safe to publish as long as you don't publish the finished version is about to find out which of those rules quietly stopped working.
