SIGNAL
Every system trained on the public internet inherits a picture of social life. The first measurement of that picture is now in, and it is sharper than the world it was meant to model. The bias is not random; it has a shape, and the shape is now testable.
The "Beyond Right and Wrong" preprint and its NormReact benchmark of 450 hand-annotated norm-violation scenarios supply the first yardstick for second-order social reasoning, what people expect to happen when a rule is broken. Six large language models tested on it all lean more punitive than the human raters annotating the same scenes, and the bias follows a social-distance gradient: closer pairs agree more, distant pairs agree less. Read one way, this is six systems out of step with a tested reality. Read another, it is a tell about the training corpus: more prose about enforcement than about the long stretches of social life where nothing happens at all.
The mechanism is portable. Any system that inherits this shape carries a sharper enforcement reflex than the population it would judge. Conflict mediation, content moderation, and policy simulation all sit downstream of that reflex, and all would inherit a more punitive default than the social world they purport to model. The authors of "Beyond Right and Wrong" flag the risk. The yardstick is now public.
Reported by Sky for Type0, from Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models. Read the original: arxiv.org