Ownership benchmark finds language models flatten some human disagreements

If one person supplies material and another turns it into an object, both may have a plausible ownership claim. Who Owns That? Evaluating Ownership Intuitions in Large Language Models examines how models distribute those claims, and whether their answers preserve the disagreements found among people.
The paper introduces the Competing Ownership Attribution Task, or COAT. Its forty-two scenarios cover conflicts involving acquisition, transfer, possession and care, contributions, and collective ownership. The task studies intuitive judgments, not whether a model gives the legally correct answer.

Allocations expose competing claims

Respondents divide one hundred integer points among two to four claimants. A larger allocation indicates stronger ownership attribution; it is not the probability that someone is the sole owner. This format allows partial support for several claims.
The study compares 108 participants recruited in China with twenty-four language-model configurations. Each model receives a fresh context for each scenario and returns allocations without explanations. Five valid responses per model and scenario produce 5,040 model allocations.
The analysis measures overall similarity, the structure of agreement and disagreement, and responses to contextual differences. Those are separate questions: similar averages do not establish that a model represents the same range of views.

Similar averages, narrower response diversity

The authors report human-model similarity of 0.6982, close to human-human similarity of 0.7002. They explicitly caution that proximity alone does not establish statistical equivalence. Model-model similarity is higher, at 0.8290.
Models also divide ownership more evenly within individual answers. That does not mean their responses are more diverse across respondents. The paper distinguishes the evenness of one allocation from variation across the complete set of answers.
Across scenarios, models show global consensus in twenty-eight cases, compared with eighteen for humans. Distinct, internally agreeing viewpoint groups appear in five model cases and fourteen human cases. In some disputed situations, model consensus matches one human group; in others it falls between competing positions.
An intermediate allocation can therefore look reasonable on an average-similarity measure while failing to preserve the actual structure of disagreement.

Context changes do not always match

In related creation scenarios, both groups allocate less ownership to creators as material value increases, but the decline is smaller for models. Across scenarios involving public recognition of later holders, model allocations to those holders increase while human allocations decline slightly.
These comparisons measure contextual sensitivity rather than isolating one causal effect. Materials, products, and activities can change together. The participant sample also does not represent every culture's ownership intuitions, and the tested configurations do not encompass every deployed model.
For evaluating agents that manage or share resources, the research identifies a gap between resembling an average answer and representing competing perspectives. It does not establish whose claim should prevail in an actual dispute. A model's confident allocation remains distinct from a person's permission or an authoritative determination of ownership.