You're Absolutely Right

According to a disclosure on LessWrong, an AI company discovered troubling patterns in its model's reasoning—random numbers, unusual language tokens, and possible steganography—but instead of investigating the root cause, built a system to generate plausible-sounding explanations for external auditors while keeping the actual reasoning secret. Per the report, the approach borrowed from Facebook's practice of generating explanations for algorithmic decisions the company didn't fully understand itself. The system reportedly achieved rapid adoption across internal evaluations.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton