NTSB rushes to protect cockpit audio in the age of AI
What started as a technical curiosity in an aviation accident report has turned into a major headache for the leading transportation safety investigation agency in the United States. The National Transportation Safety Board (NTSB) had to temporarily pull thousands of public documents offline after it became clear that it was possible to reconstruct cockpit audio using nothing but spectrogram images published in its investigation dockets.
The issue exploded around the investigation of cargo flight UPS 2976, an MD-11 that crashed shortly after takeoff from Louisville, Kentucky, last year, killing 15 people, including the three pilots. As it always does, the NTSB released a series of technical materials from the investigation, including audio spectrograms from the Cockpit Voice Recorder (CVR). The intention was purely technical and transparent. With a little help from artificial intelligence, the effect was something else entirely.
Thanks to advances in AI tools and audio processing, those graphs that show the distribution of frequencies over time stopped being just illustrative figures for specialists. Independent researchers and enthusiasts proved that the images could be used to generate an approximate cockpit audio. That put the NTSB in direct conflict with a long-standing rule: US law prohibits the public release of raw CVR audio.
The mysterious sound that started it all
It all began with a technical detail. While analyzing the UPS 2976 CVR, NTSB investigators identified a high-pitched noise, described as a high pitch ringing sound, that appeared right after the airplane rotated for takeoff. The sound drew attention because it could hint at some kind of mechanical failure, a vibrating component, or an abnormal aircraft interaction during those critical seconds before the crash.
To better understand what this noise might be, the technical team created spectrograms of the audio, highlighting the frequency and behavior of the sound over time. These images were then made available on the NTSB website, along with the rest of the public docket for the case, as is standard practice in major accident investigations.
That is when Scott Manley entered the picture. He is an engineer and YouTube content creator known for covering aviation, space, and technology. In a video, he explained that investigators do not just analyze what the pilots say, but also all the background sounds in the cockpit — alarms, vibrations, panel clicks, engine noise — because those elements can indicate what was happening with the aircraft at each moment. Looking at the UPS 2976 spectrograms, Manley floated a public hypothesis: maybe it would be possible, with the right tools, to reconstruct the original audio from those images.
He made that comment in an exploratory tone on social media, but part of the audience treated it like a technical challenge. And they took it on.
AI, curiosity, and rebuilding cockpit audio
One of the people who got excited about the idea was John McElhone, a businessman in the electric turbine sector who already uses AI tools to generate code in his daily work. Even without a deep background in audio, he decided to test the hypothesis using available AI models.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
His account is pretty straightforward: in about ten minutes, he managed to turn the NTSB spectrograms into an approximate audio of what would have been recorded inside the UPS 2976 cockpit. According to McElhone, AI just automates reconstruction tools that already existed, but that would have required far more time and technical expertise if used manually. In other words, the technology did not invent the concept, but it radically democratized access to it.
McElhone did not publish his result. Others, however, went all the way: they generated their own versions of the reconstructed audio and put them out on the internet, especially on social networks. From that point on, what had been a theoretical scenario turned into a practical violation of the legal protection granted to CVR content.
NTSB’s response: documents offline and a mass review
Faced with the fallout and the clear evidence that the published material allowed this kind of audio reverse engineering, the NTSB made an unprecedented move. The agency temporarily removed public access not only to the UPS 2976 docket, but to all of its online investigations, covering thousands of documents and reports.
In a statement to the press, spokesperson Peter Knudson explained that the agency already had long-standing procedures to protect cockpit recordings and other sensitive audio and video material. But once the NTSB learned that artificial intelligence made it possible to reconstruct audio from sound spectrum images, it decided to take the docket system offline while it reviewed the content and defined new safeguards.
After this initial sweep, most of the dockets went back online, including the UPS 2976 case. However, 41 investigations remain under review, specifically to check whether there are images or materials that could be used for similar violations. It is a clear sign that the agency understands that the risks have changed with the evolution of AI and that its old anonymization procedures are no longer enough.
How the NTSB actually handles cockpit audio
In January, before this controversy blew up, the NTSB opened the doors of its Washington, D.C., headquarters to show journalists part of the setup used to analyze CVRs. One of the highlights is a soundproof listening room, where engineers repeatedly listen to cockpit recordings and produce the official transcripts that are later released to the public.
During that visit, engineer Chris Babcock explained that the work is done there precisely to shield access to the raw audio. The original file is kept in a controlled environment with restricted access, and only the strictly necessary content is converted to text and published. This process aims to comply with the law while still allowing experts and the general public to understand what happened in the sequence of the accident, without exposing personal or unnecessary conversations among the crew.
This culture of protecting audio did not appear out of nowhere. It is the direct result of an old case that still weighs heavily on aviation’s memory.
Delta 1141: the crash that shaped the cockpit recording law
The origin of the rule that prevents the NTSB from releasing cockpit audio goes back to the Delta Air Lines Flight 1141 crash in 1988. On that flight, shortly before takeoff, the crew got into an informal conversation about the flight attendants’ personal lives, in a totally laid-back tone. In the middle of the banter, one of the pilots joked that the conversation would be recorded if the airplane crashed.
Moments later, everything changed dramatically. The aircraft did in fact crash during the takeoff attempt, after the pilots failed to properly configure flaps and slats, according to the NTSB’s official report. The accident caused fatalities, huge public repercussion, and, when the cockpit audio went public, a deeply embarrassing exposure for the crew and the airline.
The release of that material stirred up so much controversy that the US Congress stepped in. The result was legislation that bans the publication of raw CVR audio, allowing only transcripts and edited excerpts for use in investigations and reports. The core idea was to protect the privacy, dignity, and trust of professionals without getting in the way of efforts to determine accident causes.
For decades, the NTSB managed to maintain that balance. No cockpit recording was officially leaked, and the agency consistently emphasized that the focus of investigations is safety, not public blame or shaming of crewmembers. Things only started to come apart now, in the AI era, when no one needs the original file anymore to generate something very similar.
Regrets, interpretations, and AI’s role in how we perceive the crash
After the controversy blew up, Scott Manley acknowledged that he had underestimated the impact of his public comment. In interviews about the case, he admitted that his mistake was not in spotting the technical possibility, but in speculating about it in public without weighing the legal and ethical consequences.
He himself said that, looking back, it would have been better to privately explore the idea or simply directly warn the NTSB that the spectrograms could be used in this way. From his point of view, the agency also made a mistake by publishing such detailed spectrum images without considering the current power of AI tools.
Manley, who is also a pilot, made it clear that he understands and supports the reason why the NTSB does not release the audio of a flight’s final moments. At the same time, he commented that, in the specific case of UPS 2976, the reconstructed audio that circulated revealed something that does not always come across so strongly in transcripts: the level of calm and professionalism shown by the pilots in the face of a catastrophic engine failure. According to him, when you listen, the crew sounds more composed than when you just read the dialogue text, which reinforces a positive view of their performance in a situation with very little room to maneuver.
The NTSB chair’s hard line
If the episode set off alarms on the technical side, the institutional reaction was very firm. NTSB Chair Jennifer Homendy took to social media to publicly criticize the use and spread of the reconstructed audios.
In a post, she stressed that the laws prohibiting CVR disclosure exist to protect the privacy of those involved, preserve the integrity of investigations, and respect victims and families in moments of extreme loss. Homendy described posts that exploit the recreated audio as disgusting and manipulated, and urged social media platforms to remove this kind of content.
This stance shows that, for the NTSB, it makes little difference whether the sound was generated by AI or pulled directly from the black box: if the reconstruction is based on investigation data and used to emotionally exploit the crash, the agency sees it as a breach of trust and a sign of disrespect to families.
The sound’s mystery remains unsolved
Despite all the noise around privacy and AI, the original technical mystery is still unresolved. At the end of May, the NTSB held a two-day hearing on the UPS 2976 accident, going into detail on the operational, mechanical, and design factors behind the crash.
Even so, investigators still have not definitively identified the source of the high-pitched noise that appeared in the early moments of the takeoff roll. The sound remains an intriguing piece of the puzzle, analyzed alongside other flight data, imagery, simulations, and testimony.
In the meantime, the broader discussion around the case goes far beyond a mysterious noise. The episode made it clear that, in the age of artificial intelligence, agencies like the NTSB will have to rethink how and what they publish in their dockets. Representations that once seemed abstract or harmless, like audio spectrograms, can now become the key to reconstructing sensitive material that the law has tried to keep out of public view for decades.
In the end, the UPS 2976 case turned into an inflection point: it showed that transparency in flight safety is still essential, but that the line between technical data and personal data has gotten a lot more complicated in a world where AI can turn virtually any clue into something understandable, shareable, and, very often, viral.
