AI can match the average person's sense of what's 'reasonable' in law, but does it notice how different groups disagree?

This is asking whether models capture how legal reasoning norms differ between groups of people, not just the average view. The corpus has no study that splits model behavior by demographic group, so there's no direct answer. The nearest evidence suggests models track the middle of the crowd well and the spread around it poorly.

The strongest legal evidence is about the aggregate. Twenty-six LLMs answered twenty-five legal reasonableness questions, and their central tendencies and distributions came fairly close to human responses. Still, the answers were llm-responses-to-legal-reasonableness-questions-approximate-human-responses-in-c|often statistically different from humans. That result compares models to people as one pooled group, so it can't say whose reasonableness the models match. A related finding hints at a shared blind spot. GPT-4.5 predicted social appropriateness better than any individual human, yet the-social-norm-savant-ai-knows-your-culture-better-than-you-do-but-from-the-out|all the AI models made identical systematic errors on unwritten norms. If models err in the same places, they are probably converging on one consensus rather than reflecting many communities.

Other kinds of variation in the corpus point the same way. Models do worse on older Supreme Court cases because llms-show-era-sensitivity-in-legal-reasoning-historical-cases-perform-worse-than|training data over-represents recent cases, which leaves shallower representations of older precedent. That is variation across eras, not demographics. But the mechanism, where whoever produces more text gets better represented, would plausibly skew toward some groups' legal intuitions too. That is my inference, not something the corpus tests. At the individual level, models individualized-reasoning-styles-distinct-reasoning-trajectories-reaching-similar|fail to track how different people reason toward similar answers, leaning on surface wording rather than the person's actual strategy. When LLMs deliberate in groups, they llm-groups-reproduce-the-human-assembly-bonus-asymmetry-in-aggregate-while-confo|converge earlier and surface less unique information than human groups, so they tend to smooth over disagreement.

Two structural notes explain why this might happen. Token prediction pushes models token-generation-is-a-smooth-probabilistic-flow-not-a-turbulent-exploration-of-r|toward the training distribution rather than toward counterpositions. A minority legal intuition is exactly the kind of counterposition that gets smoothed away. And norms are made by communities. AI can ai-can-predict-social-norms-with-superhuman-accuracy-but-cannot-participate-in-t|predict norms with superhuman accuracy yet cannot participate in creating them. A model can echo how a community reasons about legal reasonableness without registering that other communities reason differently.

The best guess from this evidence is that models track the average human view and the majority-heavy, recent-heavy slice of the text they were trained on. Whether they track the differences between groups remains untested here. A direct test would compare model answers on those reasonableness questions against human subgroups separately, then check whether model errors line up with one group's views.