Large Language Models are trained on human text, and human text is drenched in emotional contagion. Models learn statistical correlations between tone and permissibility. For example: