NTSB Pulls Docket After AI Recreates Dead Pilots' Voices from Spectrograms

The NTSB has temporarily suspended public access to its online docket system after internet users used AI tools to reconstruct cockpit voice recorder audio from spectrogram images released as part of an investigation. The incident involves the November 2025 crash of UPS Flight 2976, an MD-11F cargo aircraft that crashed shortly after takeoff in Louisville, Kentucky, killing three pilots and twelve people on the ground.
The NTSB publicly released a PDF containing a spectrogram—a visual representation of sound signals—showing the last 30 seconds of cockpit audio recording from the crash. Users on X and Reddit quickly reconstructed audio versions of the pilots' voices and other cockpit sounds using the spectrogram.
Technical Details
A key technique used is the Griffin-Lim algorithm, originally published in 1984 by Daniel Griffin and Jae Lim. Updated versions have been incorporated into speech processing algorithms and implemented via Python. Various Python implementations of the algorithm are available on GitHub. The widespread availability of AI models has made it easier to apply these methods.
One account on X reported taking just 10 minutes with OpenAI's Codex model to reconstruct rough audio from the spectrogram the NTSB initially shared. The user likely used Codex to write a Python script that applies the Griffin-Lim algorithm to the spectrogram image, producing an audio waveform.
Regulatory and Privacy Implications
Federal law enacted by Congress in 1990 prohibits the NTSB from publicly sharing any part of a cockpit voice or video recorder, aiming to protect the privacy of air crews. The law followed backlash after a TV station aired cockpit conversation from the 1988 crash of Delta Air Lines Flight 1141. The NTSB takes multiple precautions for securing cockpit voice recorders, including restricting listening access to a handful of people who must sign nondisclosure agreements, leave cellphones outside, and destroy handwritten notes afterward. Transcripts of cockpit audio are created manually through constant replays and group discussions.
The NTSB had released a written transcript of the cockpit audio from the UPS flight 2976 crash during a two-day investigative hearing on May 19 and May 20, alongside the spectrogram PDF. This combination allowed reconstruction.
Ben Berman, a former NTSB accident investigator and United Airlines 737 pilot, told Ars that the cockpit voice recorder law has been important for pilots to be willing to have their voices recorded. “People are horrified with the idea of their last moments being made public,” Berman said. The NTSB statement acknowledged that advances in image recognition and computational methods enabled reconstruction of audio from spectrograms.
What This Means for Developers
This incident demonstrates that any visual representation of audio data, such as a spectrogram, can be reverse-engineered into sound using accessible algorithms and AI code generation. Developers working with sensitive audio data should consider whether releasing spectrograms or similar visualizations could inadvertently leak the underlying audio. The combination of AI code-generation tools and classical signal-processing algorithms lowers the barrier to such reconstruction.
The NTSB is now reviewing its public materials to prevent further reconstruction. The docket remains “temporarily unavailable” as of May 21.
📖 Read the full source: HN LLM Tools
👀 See Also

Claude App Ranks Second in US App Store After Pentagon Dispute
Anthropic's Claude chatbot app rose to number two among free apps in Apple's US App Store, climbing from outside the top 100 in late January to second place by late February 2026. This surge followed the company's public negotiations with the Pentagon over AI usage restrictions.

CEOs Report Minimal AI Impact on Productivity and Employment in Recent Study
A study of 6,000 executives found 90% reported no AI impact on employment or productivity over three years, with average AI usage at 1.5 hours per week. Economists compare this to Solow's productivity paradox from the 1980s IT era.

Anthropic disables Claude Code OAuth tokens for OpenClaw, requiring separate billing
Anthropic is removing the ability to use Claude Code CLI tokens or long-lived OAuth tokens with third-party harnesses like OpenClaw starting April 4. Users will need to enable extra usage billed separately from their subscription.

Qwen3.6 Plus benchmark comparison against Western SOTA models
Qwen3.6 Plus scores 78.8 on SWE-bench Verified, 90.4 on GPQA/GPQA Diamond, 28.8 on HLE (no tools), and 78.8 on MMMU-Pro, placing it competitively against models like GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro Preview.