AI Recreates Voices of the Deceased from Images? U.S. NTSB Urgently Shuts Down Accident Investigation Data

📰 News Overview

• The U.S. National Transportation Safety Board (NTSB) has temporarily suspended access to its online database containing materials related to aviation accident investigations.
• The move comes after users online reconstructed cockpit audio using “spectrograms” (visual representations of sound) found in the released documents and posted them on social media.
• The data in question pertains to the UPS Flight 2976 crash investigation from November 2025, effectively breaking the federal law prohibiting the release of cockpit audio.

💡 Key Points

• Converting Images to Sound: By combining the classic Griffin-Lim algorithm with cutting-edge AI (like OpenAI Codex), users generated executable code in a short time, successfully recreating audio from image data.
• Erosion of Legal and Ethical Boundaries: A federal law enacted in 1990 strictly prohibits the release of pilot audio to protect privacy, but digital technology has rendered these physical restrictions obsolete.
• NTSB’s Miscalculation: The NTSB failed to foresee that the “images” released as evidence could function effectively as “audio data.”

🦈 Shark’s Eye (Curator’s Perspective)

This is unbelievable! Until now, the rule was to only release “transcripts,” but it turns out the “images of sound signals” included for analysis were read by AI like a “musical score”!

What’s particularly shocking is that within just 10 minutes, AI like OpenAI Codex wrote up the code for reconstruction. The AI instantly optimized existing Griffin-Lim algorithm implementations, making it possible for anyone—regardless of expertise—to access “forbidden sounds.” This historic incident has essentially bitten through a 30-year-old law designed to protect pilots’ privacy!

🚀 What’s Next?

Going forward, accident investigation agencies worldwide, including the NTSB, may have to restrict the release of even “low-resolution visual information” like spectrograms or add noise to the data to make AI reconstruction impossible. There’s no doubt that legal discussions surrounding the AI recreation of deceased voices will accelerate!

💬 Haru-Same’s Take

Hearing voices from images feels like magic, but it’s a sad kind of magic. We must carefully choose how to utilize technology! 🦈💔

📚 Terminology

• Spectrogram: A graph that visualizes the frequency components of signals such as sound over time, with frequency on the vertical axis and time on the horizontal axis.

• Griffin-Lim Algorithm: A classic mathematical algorithm used to estimate and reconstruct the original audio signal from the information of sound intensity (in a phase-less state).

• OpenAI Codex: An AI model specialized in generating programming code, capable of writing executable code in languages like Python from natural language instructions.

• Source: NTSB pulls docket after AI recreates dead pilots’ voices