AI clones dead pilots' voices from a single image, jolting safety norms
NTSB pulled its public docket offline after readers used AI to reconstruct cockpit voices from a spectrogram. India's AAIB and DGCA face the same gap.
The News
The US National Transportation Safety Board temporarily pulled its public accident docket system offline last week after members of the public used AI tools to recreate the cockpit voices of pilots killed in the UPS Flight 2976 crash at Louisville in 2025. The agency restored most files on Friday but kept 42 investigations closed for review, including UPS 2976.
The reconstruction did not begin with a sound file. The docket included only a spectrogram, a mathematical image that maps audio frequencies. People combined that image with the published transcript, then ran the pair through readily available AI tools, including OpenAI's Codex, to approximate what the pilots sounded like in the cockpit's final moments.
Scott Manley, a YouTuber better known for explaining orbital mechanics, was first to publicly flag that the loop was technically feasible. Within days, reconstructed clips were circulating on social platforms. Federal law bars the NTSB from publishing raw cockpit audio. Spectrograms were not on that list. They are now effectively treated as if they were.
Why It Matters
This is the first credible case of a major safety regulator retreating from open-data publishing because of generative AI. The NTSB docket is one of the most respected accident transparency tools in the world. Researchers, journalists and rival regulators rely on it. Pulling it, even briefly, raises uncomfortable questions about investigative transparency in a world where any published artifact can be reverse-engineered into something far more personal.
The shock recalls 2023's first wave of cloned-voice fraud, when criminals used short clips lifted from YouTube to fake ransom calls. The difference is that the source material now requires no audio at all. A static image, in the right hands, is enough.
Cockpit voice recordings are protected precisely because they capture human beings in their final minutes. If those voices can be reproduced from peripheral data, the next debate will be over how much of the docket should remain public at all.
Indian Angle
India's Aircraft Accident Investigation Bureau publishes dockets and final reports, including spectrogram-style technical appendices, with policies modelled closely on the NTSB. Following the Air India Boeing 787 crash near Ahmedabad in June 2025, the AAIB faced unprecedented scrutiny over what would be released from the cockpit voice recorder. The NTSB episode now hands AAIB and the Directorate General of Civil Aviation a fresh warning on what to redact before publication.
There is a second-order risk for Indian carriers. IndiGo's record fleet expansion and Tata's ongoing Air India rebuild mean the country's commercial aviation surface area is widening rapidly. A leak of recoverable cockpit-voice data carries reputational consequences for families, airlines and insurers, and Indian voice-AI specialists such as Sarvam and Krutrim already build models capable of cloning regional-language speech, including Hindi and Tamil chatter, with limited input.
India's Digital Personal Data Protection Act, 2023 treats voice as personal data, but posthumous rights are unresolved. Expect MeitY and the DGCA to be asked, quietly at first, whether existing notifications already cover this kind of reconstruction.
FAQ
What exactly did the NTSB pull offline?
Its accident docket system, which hosts technical files for closed investigations. Most files were restored later the same Friday, but 42 investigations remain closed for review, including UPS Flight 2976. The agency has not given a fixed date for restoring those particular files.
How was the audio reconstructed without any audio file?
Users took the publicly released spectrogram, a frequency-mapped image of the cockpit recording, and paired it with the published transcript. They fed both into generative AI tools, including OpenAI's Codex, which approximated the original voices closely enough to circulate on social platforms.
Does this affect Indian accident reports?
Not yet. India's AAIB publishes similar technical material, including for the Air India Ahmedabad crash. The DGCA has not commented, but the precedent will likely force a review of what spectrogram-level data is shared in future Indian dockets.
Is there a law against cloning a deceased person's voice in India?
Indirectly. The Digital Personal Data Protection Act, 2023 covers voice as personal data, but posthumous rights are not explicitly addressed. Family-led consent regimes exist in case law but have not been tested against this kind of AI reconstruction.
This story was reported by TechCrunch. Read the full original coverage at TechCrunch.