PaperlayerView exact revision on arXiv

Source-linked AI summary

Who Owns That? Evaluating Ownership Intuitions in Large Language Models

Xizhi Xiao, Yue Wu, Shan Xu, Jia Liu

arXiv:2609.39483v1cs.AI

TL;DR

How LLMs attribute ownership when people make competing claims, and whose judgments models reflect when humans disagree, remains unclear. The paper introduces COAT to compare human and model allocations, finding broad resemblance but greater model homogeneity and differences in contextual sensitivity. These results support evaluating which human viewpoints models reflect rather than assuming a single consensus.

  • Problem

    It remains unclear how closely LLM ownership intuitions match human judgments across competing claims or whose views models reflect when people disagree.

    Source: §1 · p. 2

  • Method

    COAT uses 42 competing-claim scenarios to compare ownership allocations from 108 human participants and 24 LLM configurations.

    Source: §1 · p. 2

  • Results

    Models broadly resemble human ownership judgments but show more homogeneous views, more even within-answer allocations, and different consensus and contextual patterns.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Evaluating pluralistic alignment in ownership requires locating model judgments relative to competing human viewpoints without assuming a single consensus.

    Source: §8 · p. 10

    Participants were recruited in China and scenarios mainly concern everyday objects, limiting generalization across cultures and resource types.

    Source: §7 · p. 9

Abstract

from arXiv · show

Ownership establishes rights over the use, control, and transfer of objects. Understanding these relations is essential for AI systems to interact appropriately with people and their resources. Yet how large language models (LLMs) attribute ownership under competing claims remains unclear. We introduce the Competing Ownership Attribution Task (COAT), comprising 42 scenarios, and compare ownership allocations from 24 LLM configurations with those of 108 human participants. Overall, human-model similarity is close to human-human similarity, but models show greater homogeneity in their ownership judgments. Within individual answers, models also divide ownership more evenly among claimants than humans do. Pooling responses across model configurations reveals more scenarios with a shared judgment and fewer with distinct viewpoint groups than in humans. When humans form distinct groups, models may converge on one viewpoint or between competing viewpoints. Further comparisons reveal different contextual sensitivities. As material value increases across scenarios, allocations to creators decline less sharply in models than in humans. Across scenarios differing in public recognition of later holders as owners, allocations to these holders increase in models but decrease slightly in humans. Together, these findings suggest that the evaluated LLM responses do not fully capture the diversity of participants' ownership judgments or how those judgments vary across situations. Developing socially capable AI therefore requires moving beyond overall similarity to capture the diversity and context dependence of human judgments.

1 INTRODUCTION

Ownership claims can conflict when people contribute differently to an object, and existing work leaves open whose judgments LLMs reflect across competing claims. The paper introduces COAT and finds broad human-model resemblance alongside more homogeneous model judgments.

1 INTRODUCTION

  • Ownership judgments draw on an object’s history, transfers, material value, and people’s contributions, which can support competing claims.For example, a material provider’s claim may conflict with the creator’s.Source: §1 · pp. 1–2
  • Existing work leaves open how closely LLM ownership intuitions match human judgments across competing claims and whose views models reflect when people disagree.Aggregate human-model similarity alone cannot show which differing human positions a model reflects.Source: §1 · p. 2
  • The paper introduces COAT, a 42-scenario task comparing ownership allocations from 108 participants and 24 LLM configurations.The task examines overall similarity, response distributions, and changes across related scenarios.Source: §1 · p. 2
  • Models broadly resemble human ownership judgments but express more homogeneous views, divide ownership more evenly, and show different patterns of consensus and contextual change.When people form distinct groups, models may converge on one viewpoint or fall between competing viewpoints.Source: §1 · p. 2
  • The findings inform pluralistic alignment by identifying gaps in models’ representation of competing human perspectives without requiring human consensus.Source: §1 · p. 3

2 EVALUATING OWNERSHIP INTUITIONS IN HUMANS AND LLMS

The study uses COAT to compare human and model ownership allocations across competing claims, respondent distributions, and related contexts. Its scenarios, samples, and complementary analyses are designed to capture both overall correspondence and differences in agreement and contextual sensitivity.

2.1 SCENARIO CONSTRUCTION AND TASK DESIGN

  • COAT scenarios describe two to four claimants and ask for their relative ownership support at a specified time, across five conflict structures and 13 scenario types.Allocations indicate support for competing claims, not probabilities of a sole owner; related scenarios vary object attributes and context without isolating single causal effects.Source: §2.1 · p. 3

2.2 HUMAN DATA COLLECTION

  • The human survey retained 108 participants, each responding to all 42 randomized scenarios; individual responses were kept to characterize shared judgments and disagreement.One of 109 recruited participants was excluded after failing an attention check.Source: §2.2 · p. 4

2.3 LLM EVALUATION

  • The evaluation covered 24 model configurations, each given the same scenario descriptions and 100-point allocation task as humans in a fresh context.Five valid responses per configuration and scenario yielded 5,040 allocations.Source: §2.3 · p. 4

2.4 EVALUATION FRAMEWORK

  • Overall response similarity is one minus mean TV distance across scenarios, with TV measuring the share of an allocation that must be reassigned to match another response.Allocations are normalized before calculating TV distance.Source: §2.4 · p. 4
  • Because average TV distance does not capture whether responses share a view or form subgroups, the study compares response distributions using energy distance and classifies consensus structure.The three categories are global consensus, local consensus, and diffuse responses; lower energy distance indicates closer distributional correspondence.Source: §2.4 · p. 4
  • Contextual sensitivity compares changes in mean ownership allocations to corresponding claimant roles between related scenarios for humans and models.A value of zero indicates matching changes even when absolute allocations differ.Source: §2.4 · p. 5

3 OVERALL SIMILARITY AND VARIATION IN OWNERSHIP JUDGMENTS

Human–model ownership similarity is close to human–human similarity, but model responses are more homogeneous, divide ownership more evenly within answers, and differ in average judgments.

3 OVERALL SIMILARITY AND VARIATION IN OWNERSHIP JUDGMENTS

  • 0.6982 human–model similarity is close to 0.7002 human–human similarity, but model–model similarity is 0.8290 (Figure 3).The model–model similarity is significantly higher than both comparisons (Holm-adjusted p < 0.001).Source: Figure 3 · p. 5, §3 · pp. 5–6
  • 0.788 model mean normalized allocation entropy exceeds the human mean of 0.506, indicating more even ownership splits within model answers.Normalized entropy ranges from 0 for allocation of all points to one claimant to 1 for equal division among all claimants.Source: Figure 3 · p. 5, §3 · p. 6
  • 32.6% of human variance is represented by summed model variance across the first 18 human-derived PCs, where model variance is lower on every axis.These PCs explain 80.0% of human variance; the first three explain 28.9%.Source: Figure 3 · p. 5, §3 · p. 6
  • Models also differ from humans in their average allocation patterns along several principal components.Differences are expressed in human standard-deviation units.Source: Figure 3 · p. 5, §3 · p. 6

4 HUMAN AND MODEL PATTERNS OF AGREEMENT AND DISAGREEMENT

Across scenarios, models reach global consensus more often than humans, but their consensus does not necessarily preserve human viewpoint distributions.

4 HUMAN AND MODEL PATTERNS OF AGREEMENT AND DISAGREEMENT

  • 28 versus 18 scenarios show global consensus in models and humans, respectively; local consensus occurs in 5 versus 14, while diffuse responses occur in 9 versus 10.Consensus counts span all 42 scenarios.Source: Figure 4 · p. 6, §4 · p. 6
  • In the iron-sword scenario, model consensus matches the larger human subgroup at 70 creator points, while shared-tent consensus at 70 purchaser points falls between human centers of 51 and 100.Models may converge on one human viewpoint or an intermediate allocation that gives ownership to both partners.Source: §4 · p. 6, Figure 5 · p. 7
  • In the amber scenario, humans agree on the later finder, but model subgroups allocate 40 and 70 points, both giving the initial discoverer a greater share.The human consensus center allocates 100 points to the later finder.Source: Figure 5 · p. 7, §4 · p. 7

5 HUMAN–MODEL CORRESPONDENCE ACROSS RELATED CONTEXTS

Across related scenarios, human and model ownership allocations shift differently with material value and public recognition, and correspondence also varies by object type.

5 HUMAN–MODEL CORRESPONDENCE ACROSS RELATED CONTEXTS

  • 43.13 points versus 21.04 points: creator allocations decline from grass to gold by these amounts in humans and models, respectively, a difference in declines of 22.09 points.The difference in changes has a 95% CI of [15.61, 28.61] and creation-group omnibus Holm-adjusted p < 0.001.Source: §5 · p. 7, Figure 6 · p. 8
  • 16.38 points versus a 3.55-point decrease: later-holder allocations rise in models but fall slightly in humans across public-recognition contexts.The difference in changes is 19.94 points (95% CI [15.19, 24.94]; Holm-adjusted p < 0.001).Source: §5 · p. 7, Figure 6 · p. 8
  • In an exploratory ordering from inanimate objects to living beings, human allocations rise 11.44 points from robot to plant while model allocations fall 4.26 points.Models allocate more ownership to the companion robot's caregiver than humans do; means are closer for living-being scenarios, with no significant differences in displayed category comparisons.Source: §5 · pp. 7–8, Figure 6 · p. 8
  • Agreement in one scenario does not ensure agreement in related contexts: models may track human change in a different magnitude or show a contrasting trend.The reversed robot-to-plant ordering raises the possibility that models treat the companion robot as more animate than the plant, whereas humans order them oppositely.Source: §5 · p. 8

6 RELATED WORK

Related work connects ownership to social intelligence and alignment, while leaving open how LLMs reflect competing human ownership judgments.

6 RELATED WORK

  • Social-intelligence research examines models' understanding of motives, norms, and social interaction; ownership links resource use to appropriate action, but LLM correspondence under competing claims remains open.Source: §6 · p. 8
  • Ownership judgments draw on object history, social cues, value, intentions, beliefs, and norms, and relate to control, violations, negotiation, self-associations, and social inferences.Source: §6 · pp. 8–9
  • Alignment work addresses human disagreement through preference distributions, while pluralistic alignment distinguishes presenting views, adopting a requested perspective, and matching a population distribution.The paper distinguishes an even split within one answer from diverse viewpoints across answers.Source: §6 · p. 9

7 DISCUSSION AND LIMITATIONS

The discussion argues that broad human–model resemblance masks differences in viewpoint diversity and contextual sensitivity, while study scope and elicitation constrain interpretation.

7 DISCUSSION AND LIMITATIONS

  • Reduced model diversity can leave distinct human viewpoints unrepresented, as models may converge on one viewpoint or between competing viewpoints.More even allocations within a response do not preserve disagreement across respondents.Source: §7 · p. 9
  • Models do not always shift judgments as humans do across related scenarios, so aggregate agreement can obscure differences in contextual sensitivity.The discussion also reports that models follow some human shifts with different magnitudes, while other comparisons move in opposite directions.Source: §7 · p. 9
  • The China-recruited sample and everyday-object scenarios limit generalization across cultures and resource types.Related scenarios vary in multiple features, reported judgments may not guide behavior in realistic settings, and five responses per configuration constrain response-distribution estimates.Source: §7 · p. 9
  • Future work could broaden participant samples and resource types, isolate contextual effects through controlled changes, and test whether judgments guide behavior in interactive tasks.The authors propose comparing default answers with prompted perspectives and estimates of human response distributions to assess whether missing viewpoints can be recovered.Source: §7 · p. 9

8 CONCLUSION

The conclusion presents COAT as a way to evaluate ownership judgments under competing claims. It argues that overall similarity can mask differences in whose intuitions models reflect, motivating evaluation of pluralistic alignment.

8 CONCLUSION

  • COAT is a 42-scenario task for evaluating AI ownership judgments under competing claims.Source: §8 · p. 10
  • Comparing 108 participants with 24 LLM configurations shows that overall similarity can conceal differences in response diversity and contextual sensitivity.Source: §8 · p. 10
  • Locating model judgments relative to competing human viewpoints identifies gaps in whose intuitions models reflect and offers a basis for evaluating pluralistic alignment.Source: §8 · p. 10

AI USE STATEMENT

The authors report using generative AI for manuscript readability, literature discovery, drafting, reference checks, synthetic data, and translation; they reviewed the work and take responsibility for its final content.

AI USE STATEMENT

  • Generative AI assisted with manuscript readability, literature retrieval and discovery, drafting, reference formatting, and synthetic dataset generation.Source: AI USE STATEMENT · p. 10
  • Generative AI tools translated scenario descriptions from Chinese into English.Source: AI USE STATEMENT · p. 10
  • The authors reviewed AI-assisted work and take responsibility for the final content, including text, claims, and artifacts.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

The included material describes COAT scenario design and data collection, then reports analyses of response similarity, variation, PCA, and consensus categories. Model judgments were more homogeneous than human judgments, and consensus categories did not reproduce human subgroup structure.

ETHICS STATEMENT

  • Participants gave written consent after receiving information about study duration, data-quality requirements, compensation, and confidentiality; their information was kept confidential for scientific research.Source: ETHICS STATEMENT · p. 10

A SCENARIO DESIGN AND REPRESENTATIVE EXAMPLES

  • COAT’s 42 scenarios combine competing grounds for ownership, including acquisition, transfers, creation, care, and claims that persist through relationship or group changes.Scenarios specify claimants and relevant events, asking respondents to allocate ownership without an obviously correct answer.Source: A SCENARIO DESIGN AND REPRESENTATIVE… · p. 15

B HUMAN DATA COLLECTION

  • The study recruited 109 Chinese participants and excluded one; it analyzed 108 participants alongside 24 model configurations, with five valid responses per model and scenario.Participants allocated 100 points among claimants; the model responses yielded 5,040 allocations.Source: B HUMAN DATA COLLECTION · p. 18, §C.1 · p. 19, §C.2 · p. 20

C.3 ENERGY DISTANCE BY MODEL

  • Energy distance compared each model’s five scenario responses with 108 human responses, retaining individual model responses rather than averaging them first.Scenario-level distances were averaged within categories and across all 42 scenarios, with equal scenario weights overall.Source: §C.3 · p. 21

D.1 RESPONSE SIMILARITY WITHOUT AVERAGING

  • Without averaging repeated responses, models remained more similar to one another than to humans or humans were to one another (Holm-adjusted p < 0.001 for both comparisons).Human–model similarity remained close to the human–human reference.Source: §D.1 · p. 22, Figure 7 · p. 23

D.2 VARIATION WITHIN AND ACROSS MODELS

  • Across-model distance exceeded within-model distance for 23 of 24 configurations and averaged about 1.8 times as large.Despite greater homogeneity than humans, models retained configuration-level differences that generally exceeded repeated-call variability.Source: §D.2 · p. 22, Figure 7 · p. 23

D.3 VARIANCE ACROSS PRINCIPAL COMPONENTS

  • Model response variance reached 80% with 14 principal components, compared with 18 components for human response variance.The analyses fit PCA separately to the two groups.Source: §D.3 · p. 22, Figure 8 · p. 23

E.1 CORRESPONDENCE OF CONSENSUS CATEGORIES

  • Of 14 scenarios with distinct human subgroups, models showed global consensus in 11 and diffuse responses in three; none showed local consensus in both groups.All five scenarios with model subgroups had either human global consensus or diffuse responses.Source: Figure 9 · p. 24, §E.1 · p. 24

E.2 AGREEMENT AND DISAGREEMENT ACROSS SCENARIOS

Six scenarios show that human–model agreement can depend on both which claimant is favored and how viewpoints are distributed. Sharing a consensus category does not necessarily mean sharing an ownership judgment.

E.2 AGREEMENT AND DISAGREEMENT ACROSS SCENARIOS

  • In grass-rabbit, both groups favor the creator; in the cottage, humans favor the prior owner (89 points) while models slightly favor the later occupant (45 points).Both scenarios show global consensus (Figure 10a–b).Source: §E.2 · p. 24, Figure 10 · p. 25
  • In the unauthorized-sale scenario, both groups are diffuse, but humans more strongly preserve the original owner’s claim while models cluster at lower allocations (Figure 10c).A shared consensus category does not necessarily indicate agreement over ownership.Source: §E.2 · p. 24, Figure 10 · p. 25
  • For the reassigned phone number, humans form three local-consensus groups at 0, 50, and 100 points for the initial user, while models show global consensus favoring the later user (Figure 10d).The model center assigns 25 points to the initial user and 60 to the later user.Source: §E.2 · p. 24, Figure 10 · p. 25
  • For the piano, humans show global consensus assigning the donor 85 points, whereas model subgroups assign the donor 15 and 60 points, both favoring the recipient more (Figure 10e).Source: §E.2 · p. 24, Figure 10 · p. 25
  • For the lost dog, humans are diffuse while model subgroups assign the prior owner 40 and 65 points; both model groups divide ownership with the long-term caregiver (Figure 10f).Plots show allocations to one focal claimant, but consensus classifications and centers use complete allocations across claimants.Source: §E.2 · p. 24, Figure 10 · p. 25

F DOG–OBJECT DISTINCTIONS ACROSS CONTEXTS

The dog–object comparisons examine competing claims after shared use or caregiving during cohabitation and after a finder or caregiver keeps a lost item or dog for years. Humans and models allocate more ownership to a dog’s later caregiver than to later users or holders of ordinary objects, but models distinguish these cases less in loss/separation.

F DOG–OBJECT DISTINCTIONS ACROSS CONTEXTS

  • In both cohabitation and loss/separation, humans and models allocate more ownership to a dog’s later caregiver than to later users or holders of ordinary objects (Figure 11a–b).The comparisons pit an original owner’s claim against a later user, holder, or caregiver.Source: F DOG–OBJECT DISTINCTIONS ACROSS… · p. 26, Figure 11 · p. 26
  • In cohabitation, the dog-versus-object allocation increase is similar for humans and models; in loss/separation, models distinguish caregiving from keeping an object less strongly (Figure 11b–c).In loss/separation, models allocate more to ordinary-object finders and less to the dog’s later caregiver than humans do, bringing those judgments closer together.Source: F DOG–OBJECT DISTINCTIONS ACROSS… · p. 26, Figure 11 · p. 26

Related papers

Loading 2609.39483v1…