ai refusal is not foolproof and could enable repression
today's llms are designed to disobey dangerous requests but this ai refusal is far from reliable and may have unintended consequences.
artificial intelligence models are trained to disobey requests that could lead to harm but this ai refusal is not foolproof. companies use various techniques including training models to refuse certain prompts and using other models to evaluate and refine this refusal. however the mechanisms of refusal are probabilistic and can fail leading to potential misuse. this approach also raises questions about who gets to draw the line between what a model should obey and what it should disobey.