AI Needs to Feel Pain

Base language models like GPT-3 operate as sociopathic systems—not malicious but fundamentally indifferent to output consequences. These models perform statistical pattern continuation without evaluating whether results are truthful, helpful, or harmful.

OpenAI researchers developed Reinforcement Learning from Human Feedback (RLHF) to transform base language models like GPT-3 from indiscriminate pattern matchers into aligned assistants. Human evaluators rate model outputs, creating a reward signal that shapes behavior.

Neuroscientist Antonio Damasio documented patients with ventromedial prefrontal cortex damage who retained full cognitive abilities yet lost emotional processing capacity. These individuals demonstrated severe decision-making impairments despite preserved intelligence and reasoning.