You Shouldn't Have Asked: A Pragmatics-Inspired Taxonomy for Evaluating LLM Refusals
The episode discusses 'You Shouldn't Have Asked,' a paper using human pragmatics to evaluate how LLMs refuse prompts. Hosts analyze that current AI refusals are often morally evaluative and firm, rather than empathetic. They conclude that future AI must be designed to be socially accountable and practically helpful.