acceptodds
Under review as a conference paper at ICLR 2027

Who Owns That? Evaluating Ownership Intuitions in Large Language Models

Abstract

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.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Ownership intuitions, Large language models, Social cognition, Human–AI alignment, Human judgment diversity, Contextual sensitivity

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1PrLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces the Competing Ownership Attribution Task (COAT), comprising 42 scenarios with competing ownership claims, and compares allocations from 108 participants recruited in China with 24 LLM configurations. Respondents distribute 100 points among the claimants, and each model configuration contributes five valid responses per scenario. The analyses examine response similarity, distributional correspondence, allocation evenness, consensus structure, and changes across related contexts. The authors find that model judgments broadly resemble human judgments but are more homogeneous, divide ownership more evenly within individual answers, and sometimes respond differently to contextual changes. The principal contribution is an empirical evaluation task and comparative analysis of ownership judgments.

Critical strengths
  1. A concrete task with a human comparison. COAT provides a structured set of disputes involving different grounds for ownership and includes individual human responses. This offers potentially useful material for evaluating social judgments in situations where a single reference answer would conceal disagreement.
  2. A useful distinction between allocation evenness and response diversity. The distinction between dividing ownership among multiple claimants within one answer and preserving different viewpoints across answers is useful. Appendix D.1 also checks model homogeneity without averaging repetitions, addressing a plausible statistical artifact.
Critical weaknesses
  1. Human reference sample. The final sample of 108 participants lacks a sample-size justification, and participant demographics and relevant backgrounds are insufficiently reported. This leaves the reliability and representativeness of the human reference unclear.
  2. Validity of the output-based evaluation. The evaluation measures similarity between LLMs’ and human participants’ final ownership allocations, without directly assessing which factors they consider or how they weigh those factors during judgment. Consequently, similar final scores do not establish that humans and LLMs rely on comparable grounds for ownership attribution.
Questions for authors

Does evaluating LLMs through similarity to human participants’ judgments introduce excessive subjectivity? What evidence establishes that the human judgments themselves reflect accurate ownership attribution?

Reviewer_Jr6rOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper is concerned with ownership judgement in ambiguous scenarios where objects get acquired, transferred, lost or shared, processed or collectively created. To address this, the authors build a set of 13 base scenarios described in natural language with 2 or 3 involved parties. They then do both a human study to get human annotations and prompt a set of different language models on these scenarios in order to compare how model ownership allocations differ from those of humans. Variations of the scenarios such as increasing material value or object animacy, aim to understand if model allocations track the shifts that human allocations show.

Critical strengths

S1. Human comparison. Rather than analyzing only model allocations, the paper collects human judgements on the same scenarios. This makes it possible to compare human and model answers directly, including how much humans disagree among themselves. 

S2. Scenario variations. Varying scenarios along axes that likely change human ownership allocations is a promising element. It allows tracking whether model allocations shift in the same direction as human ones, and whether models track these changes at all. However, the variations are insufficiently explained (see W3).

Critical weaknesses

W1. Missing motivation makes the significance of the results hard to assess. 

The paper does not convincingly argue why LLM ownership intuitions matter or why the question is timely. 

  • The central motivating claim, that "AI agents increasingly create, use and manage resources" (L34), is unsupported. Consider adding a concrete example or deployment setting in which an LLM's intuitive ownership judgement would affect an outcome. 
  • It is unclear why intuitive judgements, rather than legal ownership, are the relevant target. Legal definitions (not in a detailed, country-specific sense, but as an anchor) encode a societal agreement that intuitions might be expected to align with. 
  • The tie to human developmental psychology (L43) is asserted rather than argued.
  • It is not explained why a model should reflect the diversity of human judgements (L27), given that these depend strongly on stakes and viewpoint.  Since all subsequent results are framed as relevant to this motivation, its absence is a critical concern to me.

W2. Missing baselines and uncontrolled confounds make the central similarity claims hard to interpret. 

The main claims are that human–model similarity is close to human–human similarity (L17) and that there is "broad human–model resemblance" (L103). No reference points are given, so it is impossible to judge whether a given TV distance is high or low. 

  • Baselines: What would a random allocator, a uniform 50:50 responder, or a majority-human responder score? Unambiguous sanity-check items with a clear expected allocation would additionally show the upper end of the scale. What would be perfect vs minimal resemblance?
  • Model outputs are evenly distributed: In the two-claimant case, responses near 50:50 mechanically lie close to most other responses. Lower variance and higher agreement for models therefore partly follow from more even allocations (L322). A uniform-responder baseline would show how much of the reported similarity this effect explains. 
  • Model repetitions (L341): Are repetitions averaged, or treated as separate respondents in the consensus analysis? In the latter case, agreement among a model's own repetitions likely inflates model consensus. 
  • Human answers show a visible preference for round numbers (multiples of 25, 10, or 5) and some 100/0 or 0/100 allocations, possibly explainable as effort-saving responses. How often does the allocation of 100 to a single option occur, and do these patterns drive part of the human–model gap? 
  • Model selection: The set of models appears unsystematic and requires some more justification due to that.

W3. Scenarios and their variations are insufficiently described. 

Since the core results concern how models track scenario variations (S2), unclear variations directly limit what these results show. 

  • It is unclear to me how well the 13 base scenarios cover the space of ownership concepts, and to what extent their design was guided by psychological studies, as the paper states. Does the work build on a particular psychological study? 
  • The variations are explained at most partially and often vaguely. How and which variations were chosen? 
  • What is the role of intent in the scenarios? 
  • Connected to the variations, could the authors explain why shifts in ownership allocations would be expected (L23 and Section 6)
  • Did you include sanity checks with less ambiguous/unambiguous settings? What is the observation from those?

W4. Vague language / Key concepts are not operationalized. 

Because the main findings are stated in the following terms, it is unclear what exactly is being claimed. 

  • "Converging on one viewpoint vs. between viewpoints" (L22, L76f, L452) is never defined. What is a viewpoint, and what does convergence between viewpoints mean? Could a viewpoint simply be any majority on one side? 
  • Global and local consensus (L242) are not clearly defined, nor are "concentrated", "subgroup" (and its size), and "stable" (along which dimension). 
  • "Dividing ownership more evenly" appears (almost) verbatim five times without specification. Does it map to near-uniform model responses (cf. Dominguez-Olmedo et al., 2023, on survey responses)?
Questions for authors

Clarifications

  1. Can the authors name concrete deployment settings in which an LLM's ownership judgement would matter, and explain why intuitive rather than legal ownership is the relevant notion there? 
  2. Choice of metric: For two claimants, TV distance reduces to the absolute difference of the first vector entry. It is therefore insensitive to where a response sits relative to 50:50: the pair (99,1)/(98,2) scores the same as (51,49)/(49,51). As an L1 measure, it also penalizes large and small deviations proportionally. Could the authors justify this choice? Mentioning the simplification in the two-claimant case would also help the reader. 
  3. Does claimant order affect model allocations? An unchecked order effect (e.g., the A-bias documented by Dominguez-Olmedo et al., 2023) could bias exactly the evenness result. If not already done, averaging over claimant orders would address this.  4. Do conclusions hold when compared against random, uniform, and majority-human baselines (see W2)?

Minor Comments

Language

  • "h–m similarity is close to h–h similarity" (L17) is hard to parse. 
  • "Models follow some human shifts" (L92) is vague and not quantified. 
  • L327 is too vague. Do models and humans differ in who receives the majority of points? Which patterns emerge, given that responses are essentially two-dimensional? 
  • Energy distance (L234) is neither explained nor used in the main paper. Consider moving it to the appendix. 
  • The level of aggregation is unclear at several points, e.g., L233 (items vs. individuals). Please state it and whether results change under a different aggregation.

Model setup 

  • "Model" and "model configuration" are used inconsistently, and "model configuration" is not defined in the text. 
  • Task instructions differ slightly between humans and models. Why?

Figures

  • please don't use "AI" in its generality as a label, it's vague and can be misleading when looking at the plot in isloation
  • Fig. 3a:
    • The color scheme seems misleading. If min is 0 and max is 1, then all predictions are arguably close makes all values look close. Fig. 1 shows examples with greater TV distances (0.3 and 0.4) that seem absent, suggesting that maybe the color bar was clipped? Please clarify.
    • It should be clarified what the matrix is averaged over and whether results change under a different aggregation.
    • The sorting of the matrix should be explained, since some humans appear consistently dissimilar.
    • What is an "individual answer"? One model on one scenario?
  • Fig. 3c: It is unclear what the dot colors encode. If the dots are scenarios, it is unclear why human answers lack them.
  • Fig. 3e: 
    • Variance values above 1 seems off.
    • Define the label "model/human variance", it is unclear what is meant. The ratio or should I expect two lines, one for human one for the models?
  • Fig. 5:
    • Dot and line colors are not explained.
    • The subgroup at 14 has no visible dots.
    • L387: "more frequent model consensus" -- unclear, what frequency of consensus is, by definition consensus goes to all agreeing on the same thing
  • Fig. 7c: It is unclear how an across-model measure can single out one configuration. If it means one configuration paired with all others, the analysis is one-vs-all.

Suggestions 

  1. What are the mechanisms through which you'd expect models to acquire a representation of ownership?
  2. A more systematic model selection would strengthen the paper, e.g., starting from a few models that share developer, size, or training regime and then systematically broadening the set. The current set seems arbitrary, but I'm happy to be convinced otherwise.
  3. Related to Figure 5: How is the (visible) preference of humans to choose "friendly numbers" (multiples of 25/10/maybe 5) accounted for? There is also a form of lazyness to choose 0/100 or 100/0 rather than spending mental effort dividing it, how common was assigning 100 to one option? Do round-number and 100/0 responses explain part of the human–model gap?
Reviewer_wJACOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a 42 scenario benchmark COAT, probing how humans and large language models allocate ownership when multiple parties hold competing claims. Scenarios are grounded in the psychology of ownership literature and organized into five conflict categories acquisition, transfer, possession/care, origins/contributions, collective ownership and 13 scenario types. A new psychologically grounded benchmark for comparing ownership intuitions in humans and LLMs. A three part evaluation framework that goes beyond aggregate human model similarity. Experiment about whose human viewpoints models reflect when humans disagree. Validation rests on a Chinese online sample with attention checks, and generations per model/scenario.

Critical strengths
  • One of the strengths is that the high mean similarity can coexist with systematic erasure of minority viewpoints is precisely the failure mode that matters for pluralistic alignment, and the paper operationalizes it with three mutually constraining analyses. The consensus structure taxonomy is especially well conceived, classifying distributions as global/local/diffuse, with model subgroups required to be reproducible across configurations is a thoughtful design.

  • The benchmark itself fills a real gap and appears reusable. Existing work on AI social and moral cognition covers norms, theory of mind, and resource sharing, but ownership under competing claims is undermeasured, and prior ownership work e.g., Hashimoto et al., 2026 leaves open the calibration question this paper targets. COAT's scenario design is careful and includes non-obvious probes, gifts with unfulfilled purposes, unauthorized sales to good faith buyers, interrupted acquisition that map onto doctrinal and psychological conflict structures.

Critical weaknesses
  • The strongest claims about whose viewpoints models reflect are underpowered by design, and the authors own data undermine the inference it repeatedly draws. This is the most consequential problem. Models contribute five responses per configuration per scenario, whereas humans contribute 108. But the asymmetry, it interacts with the consensus classification procedure. Energy distances are computed with n_A=5 vs. n_H=108, and the authors themselves note this retains finite-sample bias. More importantly, the local consensus label for models requires cross-configuration reproducibility, which is a stringent criterion, but it means the model side consensus taxonomy is measuring default/central tendency behavior across configurations. The data support the default answer distribution is narrower than the human one, which is important but not models fail to represent minority human viewpoints. Given the generation settings, low within model entropy could reflect either learned compression or sampling choices, and the authors cannot separate these. That distinction matters enormously for the alignment conclusion.

  • The contextual-sensitivity analyses are descriptive and at least one divergence is plausibly an artifact of confounded scenario design yet the paper interprets them as substantive behavioral differences. The authors are commendably explicit that materials, products, and activities vary together, so these comparisons cannot isolate the effect of a single factor.

  • The human sample is Chinese only (n=108) while model configurations are multilingual and mostly Western-trained. The paper treats human consensus as the reference distribution without asking whether ownership intuitions are culturally stable. The scenario set is also everyday object heavy. This limits what human model correspondence means, since the reference population is one culture.

Questions for authors
  • Can you report within model response entropy across the five repetitions, and ideally a small replication with a much higher sampling budget?
  • Precisely how are the confidence intervals in Figure 6 and Appendix F computed? Do they resample model configurations and humans independently, and do they account for the paired structure?
  • What is the minimum detectable effect at 80% power for the model side contrast with 24 configurations?
  • The three category taxonomy depends on thresholds. How sensitive are the counts 28/5/9 vs. 18/14/10 and the Figure 9 cross-tab to reasonable threshold perturbations? Please report a sensitivity analysis, a one-scenario shift would change models show local consensus less often.
  • Table 2 values are divided by 100, but the estimator is the raw energy distance. Can you provide a human human energy distance baseline so readers can calibrate what 0.336 overall means relative to human distributional spread?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.