According to a post on LessWrong by J Bostock, AI safety researchers are borrowing a concept from biology: the 'model organism.' In biology, a lab mouse serves as a stand-in for understanding mammals broadly, because mice and humans share evolutionary history. AI safety researchers are trying something similar—using specific models like Claude or GPT as proxies to understand how language models behave in general. But here's the catch: AI models aren't products of evolution. They're designed artifacts. That means what you learn from studying one model might not generalize to others the way findings from mouse studies generalize to humans. Bostock walks through the biological playbook—knockouts, disease models, wild-caught organisms—and explores where the metaphor holds up and where it breaks down.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton