The National Transportation Safety Board (NTSB) recently faced a serious privacy challenge. People used artificial intelligence to recreate the voices of pilots who died in a plane crash. This happened after the agency released investigation files that included a visual audio spectrogram.

The crash involved UPS Flight 2976, a cargo plane that went down in Louisville, Kentucky on November 4, 2025. Three crew members died, along with 12 people on the ground. The NTSB shared thousands of pages of reports and a video showing the engine separating from the wing. They also included a transcript of cockpit audio and a spectrogram image representing the cockpit voice recorder data.

While federal law bans the release of actual cockpit voice recordings, the NTSB did publish this spectrogram. A spectrogram turns sound signals into images, showing frequency and time. This image was meant for experts to study mechanical and audio clues, not to recreate voices.

How AI Recreated Pilot Voices

AI tools have advanced to the point where they can reverse-engineer spectrograms back into audio. People online used OpenAI’s Codex and other algorithms to reconstruct the last 30 seconds of cockpit audio. They combined the spectrogram with transcripts to create synthetic voices of the pilots during their final moments.

One user claimed it took just 10 minutes to produce a rough audio clip. These AI-generated recordings quickly spread on social media platforms like X and Reddit. This raised immediate privacy concerns for the victims and their families.

NTSB’s Response and Privacy Concerns

After learning about the AI recreations, the NTSB temporarily shut down its public docket system. The agency said it would review the problem and consider new ways to protect sensitive data. They restored most files later but kept 42 investigations, including the UPS crash, offline for further review.

NTSB Chairwoman Jennifer Hommendy called the situation deeply troubling. The agency balances transparency with protecting victims’ privacy. Once recreated audio hits the internet, it becomes nearly impossible to remove or control.

This incident highlights a new challenge for agencies that share technical data. Spectrograms and other visual audio forms were never expected to be turned back into sound. AI’s ability to do this forces agencies to rethink how they release investigation materials.

Experts say future policies might treat spectrograms like raw audio files, applying strict controls to prevent misuse. The NTSB is also working with social media companies to remove unauthorized AI-generated audio from public sites.

For now, this case serves as a warning. AI can bring new risks to privacy, especially when handling sensitive data about tragic events. Regulators will need to find a balance between openness and respect for those affected.