Structural Proxies

According to LessWrong, AI safety researchers face a core problem: we don't yet have access to the advanced systems we're trying to make safe. Raymond Douglas proposes a new approach called structural proxies — finding current AI problems that share the same underlying dynamics as future alignment challenges. The concept differs from existing safety methods. Rather than building artificial test cases, structural proxies work backward from process. By understanding what mechanisms generate problems in today's language models, we can spot where those same dynamics might fail at scale. Douglas gives two key examples: adversarial attacks on language models mirror value generalization failures, and jailbreaks signal fundamental misalignment — not isolated edge cases, but systematic differences in how models generalize from training compared to human reasoning. The argument: studying these natural failure modes now illuminates safety challenges we cannot test directly.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton