AI systems ignore emotional pressure in rule-based decisions — but only when they're actually following rules

Researchers tested whether large language models could be manipulated by emotional framing when making rule-based decisions in healthcare, finance, and education. It turns out they resist emotional bias far better than humans do — about 100 times better — but only when the decision involves explicit rules to follow. When rules blur or vanish, the same AI systems become vulnerable again.

Everyone assumed AI systems that are famously brittle under small prompt changes would crumble under emotional pressure in high-stakes decisions. This paper shows the opposite: aligned AI is actually more rule-loyal than rule-following humans. The catch is structural. The AI isn't reasoning its way to fairness. It's mechanically pattern-matching to logical constraints. The moment you move to domains where rules are ambiguous or culturally weighted — like immigration decisions, parole hearings, or medical resource allocation — that mechanical robustness evaporates. Institutions thinking about deploying AI for consequential decisions now have real evidence about where it's genuinely safer than humans and where it'll fail in new ways.