AI Can Reconstruct Cockpit Voice Recordings from Published Spectrograms

The NTSB confirmed in 2026 that cockpit voice recordings circulating online from the 2025 UPS Flight 2976 crash investigation were not leaked — they were reconstructed using AI tools applied to spectrogram images published in official investigation dockets.

This is a new category of problem for failure investigation: the evidence itself has not been compromised, but AI tools can now synthesize convincing audio from the visual representations of that evidence that investigators routinely make public.

How the Reconstruction Works

Cockpit voice recordings are typically not released publicly during investigations — but spectrograms, which are visual frequency-over-time representations of audio signals, are sometimes included in docket materials as analytical exhibits. Investigators have historically considered these safe to release: a spectrogram shows the structure of a sound without being the sound itself.

That assumption no longer holds. AI audio reconstruction tools can invert spectrograms with sufficient fidelity to produce intelligible speech — creating a synthesized audio file that approximates what was recorded in the cockpit, without access to the original recording.

Why This Matters for Investigations

Reconstructed audio that sounds authentic but is not can be used to advance theories about crew behavior, statements, or decision-making that may be inaccurate or misleading. In litigation contexts, such audio could be presented to juries or used to shape public narrative before facts are established.

It also creates a new challenge for investigators publishing technical materials: what was previously considered safe technical disclosure is now potentially a source of reconstructable evidence.

What Investigators and Counsel Should Know

Anyone working with NTSB docket materials — whether as investigators, expert witnesses, or counsel — should be aware that spectrogram exhibits from audio recordings may now be subject to reconstruction. The NTSB's warning signals that the agency is aware of the issue, but guidance on revised disclosure practices has not yet been issued.

This is a developing area. The intersection of AI capability and failure investigation evidence standards will require both regulatory and procedural responses in the coming years.