AI Judges People More Harshly Than Reality, Study Finds

Teaching AI right from wrong isn't enough. That's the finding of a new study that looked at how AI language models like ChatGPT understand what happens after a social rule is broken — not just whether an action is bad, but who reacts and how. The researchers built a dataset called NormReact, containing 450 everyday norm violations, such as cutting in line or breaking a promise. Each scenario was annotated by humans to show what a violator might feel and what a witness might do.

Then they tested six major AI models. The result: the AIs consistently overpredicted punishment. When a human would feel annoyed but say nothing, the AI expected confrontation or shaming. The models were also worse at reading situations involving strangers. The more distant the relationship between the wrongdoer and the observer, the more the AI assumed a harsh response. In other words, AI imagines a social world that is meaner and more punitive than the one we actually live in.

Why does this matter? AI systems are starting to play roles in conflict mediation, online community moderation, mental health support, and even policy simulations. If these systems believe that people constantly punish, shame, and retaliate, they may give advice or make decisions that feel cold, aggressive, or unfair. A chatbot helping two coworkers resolve a dispute might push for an apology or penalty too hard. A moderation tool might ban users when a human reviewer would simply scroll past.

The researchers are not saying AI is malicious. They are pointing out a blind spot: social intelligence includes knowing when not to punish, when to forgive, and how relationships shape our responses. The study gives developers a new way to test and improve these skills, so future AI can better reflect the tolerant, forgiving, and nuanced ways real communities handle conflict.