Ask an AI who owns something, and it will probably give you a tidier answer than your neighbor would.
A new paper introduces the Competing Ownership Attribution Task, a set of 42 scenarios designed to pit ownership claims against each other, say a creator versus someone who later holds or publicly claims an item. Researchers compared judgments from 24 different large language model configurations against 108 human participants on these scenarios. Overall, the AI responses matched human judgments about as closely as humans matched each other. But the models showed far less internal disagreement: where humans split into distinct camps on a scenario, models tended to converge on a single answer or land between competing views, and models divided ownership more evenly among claimants than people did.
The gap shows up most clearly in how each group handles context. As an item's monetary value rises, human allocations to the original creator drop off sharply while model allocations barely adjust. And when a later holder gains public recognition as the "real" owner, model allocations to that holder increase, while human allocations actually decrease slightly.
That is a meaningful blind spot for anything built to arbitrate disputes, write terms of service, or mediate marketplace disagreements. A model that smooths over disagreement is not neutral, it is just missing the disagreement. Average similarity scores can hide the fact that a system agrees with everyone a little and no one completely.