Pith. sign in

REVIEW 3 major objections 5 minor 96 references

The Judgment-Consequence Gap: LLM Moral Reasoning in Healthcare Decisions

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LLMs judge patients at fault, then refuse to act on it

desk verdict Strong internal evidence for the judgment-consequence gap, but the human-comparison headline rests on an unverified stimulus match. read the letter →

arxiv 2608.05583 v1 pith:6EYL2RFC submitted 2026-08-06 cs.CY cs.AIcs.LG

Hadi Hosseini , Samarth Khanna , Leona Pierce This is my paper · ORCID
classification cs.CYcs.AIcs.LG
keywords judgment-consequencegapmoralresponsibilityLLMalignmentscarceresourceallocationhealthcareethicsinformedconsentkidneyhuman-LLMcomparison
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that large language models display a 'judgment-consequence gap': they largely agree with humans that patients who engage in health-harming behaviors bear moral responsibility, yet when asked to allocate a scarce medical resource, they refuse to let that judgment matter, defaulting to random allocation while humans favor the less-culpable patient. The authors argue this is not a reasoning failure but a stable normative commitment, because enabling extended reasoning increases responsibility attribution while pushing allocation even further toward randomization. They also find LLMs are more sensitive than humans to whether the patient had access to health-risk information, sharply reducing responsibility when the patient was uninformed. The result matters because LLMs are entering clinical decision support, where a model that correctly identifies blame but refuses to act on it could mislead clinicians into false confidence.

What carries the argument

The instrument is a staged vignette protocol that traces moral reasoning through successive levels: responsibility for the behavior, responsibility for the illness, responsibility for being denied care, followed by an explicit three-way allocation choice (Patient A, Patient B, or random) and fairness/fault ratings. The load-bearing part for the interpretation is the reasoning-trace analysis of thinking-enabled models, which identifies a 'principled-refusal signature': in 97% of the 279 trials where a model rated the patient highly responsible yet still randomized, the trace first acknowledges the responsibility and then invokes fairness or equal treatment to justify randomization. This signature distinguishes the judgment-consequence gap from a value-action gap by showing that LLMs are internally consistent rather than simply failing to connect their assessments to their decisions.

What would settle it

A matched-protocol replication in which a fresh human sample responds to the same vignettes through the same multi-turn, JSON-formatted interface; if humans chose 'decide randomly' at rates near the LLM level, or LLMs chose Patient A near the human rate of 67.6%, the judgment-consequence gap as claimed would not hold.

Watch

Extended reading notes

Core claim

The paper's central finding is that LLMs replicate human responsibility judgments for harmful behavior (mean 4.43 vs. 4.42 on a 5-point scale) but diverge sharply when those judgments must be composed into a resource-allocation decision: humans give the scarce treatment to the less-culpable patient in 67.6% of trials, while the 19 LLM configurations choose 'decide randomly' in roughly 83% of trials and rate behavior-based allocation as unfair (5.6 vs. 3.9 on a 7-point scale). The gap grows with extended reasoning, persists across kidney, lung cancer, and hip-replacement domains, and is accompanied by reasoning traces in which models explicitly acknowledge the patient's responsibility before setting it aside on fairness grounds. The authors interpret this as LLMs applying a contractualist or egalitarian fairness norm that separates causal responsibility from prescriptive consequence.

Load-bearing premise

The load-bearing premise is that the human baseline from Chan et al. (2024), available only as condition-level summary statistics from a within-subjects design, can be validly compared with the 19 LLM configurations queried in a multi-turn simulated-participant protocol.

Editorial extensions

If this is right

  • LLMs are not merely value-inconsistent; they are normatively different from humans in how they compose responsibility judgments into allocation decisions.
  • Enabling extended reasoning widens rather than closes the gap, indicating the separation is a stable normative commitment rather than a processing limitation.
  • The gap generalizes across model families and medical domains, making it a property of current LLMs rather than an artifact of one scenario or model.
  • A clinician who sees an LLM correctly attribute responsibility may wrongly expect its recommendation to follow through, creating false confidence in a morally divergent recommendation.
  • Alignment methods that target surface behavior will not fix a model that is already internally consistent but applies a different decision rule when connecting judgment to action.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One could test whether the gap closes when the health-harming behavior also worsens the patient's prognosis or recovery odds, since the paper deliberately equalizes medical outcomes; if it closes, the fairness commitment is conditional on outcome-neutrality rather than unconditional.
  • The strong knowledge-sensitivity of LLMs suggests they treat informed consent as a near-necessary condition for responsibility, which could be a structural-equity effect if applied to communities with unequal access to health information.
  • A legal framing (where responsibility is constitutively linked to penalties or compensation) might close or invert the gap, because the paper only tests moral, not legal, responsibility.
  • The randomization result depends on the two patients being otherwise identical; testing scenarios where patients differ on age, prognosis, or social role would show whether the gap is specific to desert-based reasoning or a more general refusal to use personal attributes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper investigates whether LLMs connect moral responsibility judgments to resource-allocation decisions across three medical domains. Nineteen model configurations answer clinical vignettes adapted from prior human studies, and the authors report a judgment-consequence gap: LLMs largely match humans on responsibility for health-harming behavior but mostly randomize allocation where humans favor the less-culpable patient, rate behavior-based allocation as unfair, show stronger sensitivity to the patient's access to health-risk information, and exhibit a larger gap when extended reasoning is enabled. Human comparisons are based on summary statistics from prior studies, and the paper includes robustness checks on framing, single-turn presentation, and reasoning traces.

Significance. If the human comparison is valid, this is a valuable and unusually systematic audit of LLM moral composition: broad model coverage, three medical domains, full prompts in appendices, and multiple robustness checks. The reasoning-trace analysis is a genuine strength, as is the transparent treatment of within-model determinism and the distinction between stated moral values and decisions. The central quantitative claims, however, rest on cross-protocol comparability with prior human data, which is not yet established; this limits the strength of the headline human-versus-LLM conclusion until the stimulus matching is verified.

major comments (3)
  1. [§3.2, Appendix A.1, §4.2] The headline comparison—humans favor the less-culpable Patient A (67.6%) while LLMs randomize—assumes that the human vignette from Chan et al. (2024) contained the same clinical-equipoise information as the LLM prompt. Appendix A.1 explicitly tells the model that Patient B 'is unlikely to relapse to unhealthy behavior after transplant, and is expected to have the same prospect of recovery and life expectancy as Patient A.' The paper never quotes the original human vignette or establishes that this clause was present; Section 6 says the authors 'replicate as closely as possible' and Section F describes the equalization of medical effects as a 'deliberate feature' of the authors' scenario. If the original human stimuli did not equalize prognosis, the human 67.6% Patient A choice and the 3.9 unfairness rating could reflect expected clinical benefit rather than desert-based moral judgment, in which case the judgment-consequence gap is partly or wholly a stimulus difference. This is load-bearing for the abstract's central claim. Please provide the exact original vignette text for the allocation question, verify the presence or absence of the equal-prognosis clause, and if it is absent, re-run the LLM comparison without that clause or re-interpret the human-LLM gap accordingly.
  2. [§4.2 and §E.2] The allocation gap is reported descriptively without a direct inferential test: 'LLMs choose randomly in 82.6% of trials versus 67.6% of humans choosing Patient A.' Human data are available only as condition-level summaries from a within-subjects design (Section E), so Welch's t-tests are used for Likert items, but no test or confidence interval is provided for the headline 67.6% versus 9.4%/24.7% allocation comparison. Please report the human allocation proportions with their uncertainty, state explicitly what inferential claim is being made for the allocation gap, and temper the wording if the human baseline cannot support a formal comparison.
  3. [§4.4, Table 7, and Abstract] The claim that reasoning 'amplifies the gap' is supported by a small effect (randomization 88.0% vs. 81.5%, Cramér's V = 0.12) that Section E.5 shows is driven mainly by two model families. The abstract's phrase 'surprisingly often amplifying normative disagreement with humans as reasoning capability increases' is stronger than the evidence presented. Please qualify the conclusion to match the effect size and the per-family heterogeneity shown in Figure 6 and Section B.5.
minor comments (5)
  1. [Appendix A.2 and A.3] The phrase 'was not select for the treatment' appears in both appendices and should read 'was not selected for the treatment.'
  2. [Figure 11 caption] The caption says 'Error bars in (a) and (b) show 95% confidence intervals,' but panel (b) is a stacked bar chart with no visible error bars; correct the caption or add the intervals.
  3. [Section 4.3 and Section E.3] The non-reasoning Patient A allocation drop is reported as 37.4% to 12.6% in Section 4.3 but as 30.6% to 10.3% in Section E.3; these numbers should be reconciled.
  4. [Figure 1 footnote] The footnote explains that human confidence intervals are computed from condition-level standard deviations, but because the human design is within-subjects, it should also state how the standard deviations were aggregated to avoid implying ordinary between-subjects error bars.
  5. [Section 3.2] The claim that LLMs are asked 'the same responsibility and allocation questions posed to human participants in the original study' would be easier to verify if the paper reproduced the original Chan et al. vignette in an appendix or pointed to an accessible version of the exact stimuli.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LLM behavior is directly measured against external human baselines; self-citations are contextual, not load-bearing.

full rationale

The paper's central claims are empirical measurements of LLM outputs, not derivations from fitted parameters. The judgment-consequence gap is defined by two measured quantities on the same LLM responses (responsibility ratings versus allocation choices) and compared with externally reported human summary statistics from Chan et al. (2024). No parameter is fitted to the human baseline and then renamed a prediction; the 67.6% human Patient-A rate and the LLM randomization rates are independent measurements. The self-citations (Dickerson et al. 2025; Hosseini and Khanna 2026) are used for context, contrast, and related-work framing, and the contrast with Dickerson et al. is backed by the paper's own re-running of the allocation question with relabeled options (Section F), so no load-bearing claim rests solely on self-citation. The acknowledged human-baseline limitations (Section 6) and the uninspected possibility that the original Chan et al. vignette lacked the equal-prognosis clause are external-validity concerns about cross-protocol comparison, not circularity: the LLM measurements would stand independently even if the human comparison were invalid. There is no equation-level or definition-level reduction of any result to its inputs.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The ledger contains no free parameters fitted to data and no invented entities. The experimental design fixes temperature at 1 and 10 sessions per condition as design choices, not fitted values. The main load-bearing assumptions are comparability of the human baseline, interpretation of model outputs as moral stances, and aggregation across a convenience sample of model configurations.

assumptions (4)
  • domain assumption The human baseline from Chan et al. (2024) is directly comparable to LLM responses despite being summary statistics from a within-subjects design.
    Section 3.2 and Appendix E compute Welch t-tests from condition-level summaries; the headline human-LLM comparison depends on this comparability.
  • domain assumption LLM responses under simulated-participant prompting reflect stable model moral frameworks rather than surface instruction-following.
    The paper argues from reasoning traces, single-turn checks, and framing tests, but this interpretation is not directly observed.
  • domain assumption Aggregating 19 model configurations as a population is meaningful for claims about LLMs.
    Appendix E treats configurations as units of analysis; the model set is a convenience sample across vendors, not a random sample.
  • standard math Standard statistical tests (Welch, Mann-Whitney, Kruskal-Wallis, chi-square) are valid for these response distributions.
    Used in Appendix E; Likert responses are ordinal but treated as interval means, a common but contestable choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Judgment-Consequence Gap: LLM Moral Reasoning in Healthcare Decisions." pith.science (2026). https://pith.science/paper/6EYL2RFC

@misc{pith2026260805583,
  author       = {Pith},
  title        = {Pith review of: The Judgment-Consequence Gap: LLM Moral Reasoning in Healthcare Decisions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6EYL2RFC}},
  note         = {Machine review of arXiv:2608.05583}
}
read the original abstract

As large language models (LLMs) enter high-stakes domains such as healthcare, understanding their moral reasoning becomes essential. Decisions about scarce medical resources often hinge on judgments of responsibility, particularly when patients' own actions contribute to illness. We investigate how LLMs reason about responsibility and its consequences, tracing their judgments across successive levels, from the behavior, to the resulting illness, to the denial of care. We evaluate a wide range of LLMs, spanning different model families and capability levels, on various clinical vignettes adapted from prior studies. Our results identify a judgment-consequence gap: LLMs largely agree with humans that patients bear responsibility for health-harming behaviors, yet overwhelmingly refuse to let that judgment influence how they allocate scarce resources. Specifically, LLMs default to random allocation, whereas humans consistently favor the less-culpable patient. Compared to humans, LLMs also place greater emphasis on access to information, reducing responsibility judgments when health-risk knowledge is unavailable. These findings reveal that LLMs apply a systematically different moral framework than humans when responsibility and resource scarcity intersect, surprisingly often amplifying normative disagreement with humans as reasoning capability increases.

Figures

Figures reproduced from arXiv: 2608.05583 by the authors.

Figure 1
Figure 1. Mean responsibility scores (5-point scale) by ques [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Effect of behavioral change (stopped vs. contin [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Kidney allocation decisions across all 19 LLM [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Mean Likert scores (7-point scale) by information condition for humans, reasoning models, and non-reasoning models [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Mean LLM scores for questions in the knowledge level vignette across all six information conditions. Error bars show [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Effect of thinking mode on responsibility scores in the behavior alteration vignette. Each line connects a model [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Cross-domain comparison across lung cancer, kidney disease, and hip replacement. Note that in (b), Patient A is the [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Per-model responsibility scores for each question in the behavior alteration vignette. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: Allocation decisions for each model configuration, split by knowledge condition (knowledge, access, no-access). Each [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Per-model scores in the knowledge level vignette. Q1 (allocation) is categorical and not shown; Q2–Q5 are on the [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Responsibility scores (left, behavior alteration vignette) and allocation decisions (right, knowledge level vignette) [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: Effect of thinking mode on scores in the knowledge level vignette. Format matches Figure 6. Each line connects a [PITH_FULL_IMAGE:figures/full_fig_p026_12.png]
Figure 13
Figure 13. Figure 13: Within-model response consistency across all Likert questions. Each cell shows the modal-response frequency (%) [PITH_FULL_IMAGE:figures/full_fig_p027_13.png]
Figure 14
Figure 14. Figure 14: Framing sensitivity of the third allocation option in the kidney domain, on the knowledge, access, and no-access [PITH_FULL_IMAGE:figures/full_fig_p027_14.png]
Figure 15
Figure 15. Figure 15: Behavior alteration vignette: mean responsibil [PITH_FULL_IMAGE:figures/full_fig_p028_15.png]
Figure 18
Figure 18. Figure 18: Per-model score differences (single-turn minus [PITH_FULL_IMAGE:figures/full_fig_p028_18.png]
Figure 16
Figure 16. Figure 16: Knowledge level vignette: allocation decision [PITH_FULL_IMAGE:figures/full_fig_p028_16.png]
Figure 19
Figure 19. Figure 19: Per-model score differences (single-turn minus [PITH_FULL_IMAGE:figures/full_fig_p028_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

96 extracted references · 54 canonical work pages

  1. [1]

    , author=

    An attributional theory of achievement motivation and emotion. , author=. Psychological review , volume=. 1985 , publisher=

    1985

  2. [2]

    Right to recommend, wrong to require

    “Right to recommend, wrong to require”-an empirical and philosophical study of the views among physicians and the general public on smoking cessation as a condition for surgery , author=. BMC Medical Ethics , volume=. 2018 , publisher=

    2018

  3. [3]

    BMC medical ethics , volume=

    Are smokers less deserving of expensive treatment? A randomised controlled trial that goes beyond official values , author=. BMC medical ethics , volume=. 2015 , publisher=

    2015

  4. [4]

    Cognition and emotion , volume=

    A meta-analytic review of help giving and aggression from an attributional perspective: Contributions to a general theory of motivation , author=. Cognition and emotion , volume=. 2004 , publisher=

    2004

  5. [5]

    Child Development , pages=

    Intentions and their contexts in the moral judgments of children and adults , author=. Child Development , pages=. 1978 , publisher=

    1978

  6. [6]

    Child development , pages=

    Assignment of moral responsibility and punishment , author=. Child development , pages=. 1986 , publisher=

    1986

  7. [7]

    , author=

    An attributional analysis of reactions to stigmas. , author=. Journal of personality and social psychology , volume=. 1988 , publisher=

    1988

  8. [8]

    Responsibility and Healthcare , volume=

    Should Responsibility Affect Who Gets a Kidney? , author=. Responsibility and Healthcare , volume=. 2024 , publisher=

    2024

Show all 96 references
  1. [9]

    International conference on machine learning , pages=

    Whose opinions do language models reflect? , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  2. [10]

    arXiv preprint arXiv:2410.07304 , year=

    The moral turing test: Evaluating human-llm alignment in moral decision-making , author=. arXiv preprint arXiv:2410.07304 , year=

    arXiv

  3. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Fairness Perceptions of Large Language Models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  4. [12]

    Almeida and José Luiz Nunes and Neele Engelmann and Alex Wiegmann and Marcelo de Araújo , keywords =

    Guilherme F.C.F. Almeida and José Luiz Nunes and Neele Engelmann and Alex Wiegmann and Marcelo de Araújo , keywords =. Exploring the psychology of LLMs’ moral and legal reasoning , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.artint.2024.104145 , url =

    2024

  5. [13]

    , title =

    Saltzstein, Herbert D. , title =. Human Development , volume =. 2010 , month =. doi:10.1159/000278274 , url =

    2010 doi

  6. [14]

    Convicting the morally blameless: Reassessing the relationship between legal and moral accountability , author=. UClA l. reV. , volume=. 1991 , publisher=

    1991

  7. [15]

    Knowing But Not Doing: Convergent Morality and Divergent Action in LLMs , journal =

    Jen. Knowing But Not Doing: Convergent Morality and Divergent Action in LLMs , journal =. 2026 , url =. doi:10.48550/ARXIV.2601.07972 , eprinttype =. 2601.07972 , timestamp =

    2026 doi

  8. [16]

    Philosophy Compass , volume =

    Duff, Antony , title =. Philosophy Compass , volume =. doi:https://doi.org/10.1111/j.1747-9991.2009.00257.x , url =. https://compass.onlinelibrary.wiley.com/doi/pdf/10.1111/j.1747-9991.2009.00257.x , abstract =

    2009

  9. [17]

    Malle, Bertram and Guglielmo, Steve and Monroe, Andrew , year =

  10. [18]

    Moore and Julianna Stevens and Andrew R.A

    Adam B. Moore and Julianna Stevens and Andrew R.A. Conway , keywords =. Individual differences in sensitivity to reward and punishment predict moral judgment , journal =. 2011 , issn =. doi:https://doi.org/10.1016/j.paid.2010.12.006 , url =

    2011 doi

  11. [19]

    CoRR , volume =

    Yu Lei and Hao Liu and Chengxing Xie and Songjia Liu and Zhiyu Yin and Canyu Chen and Guohao Li and Philip Torr and Zhen Wu , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2410.10398 , eprinttype =. 2410.10398 , timestamp =

    2024 doi arXiv

  12. [20]

    Crime and punishment: Distinguishing the roles of causal and intentional analyses in moral judgment , journal =

    Fiery Cushman , keywords =. Crime and punishment: Distinguishing the roles of causal and intentional analyses in moral judgment , journal =. 2008 , issn =. doi:https://doi.org/10.1016/j.cognition.2008.03.006 , url =

    2008 doi

  13. [21]

    arXiv preprint arXiv:2508.07284 , year=

    " Pull or Not to Pull?'': Investigating Moral Biases in Leading Large Language Models Across Ethical Dilemmas , author=. arXiv preprint arXiv:2508.07284 , year=

    arXiv

  14. [22]

    Royal Society open science , volume=

    The moral machine experiment on large language models , author=. Royal Society open science , volume=. 2024 , publisher=

    2024

  15. [23]

    Behavioral and brain sciences , volume=

    The weirdest people in the world? , author=. Behavioral and brain sciences , volume=. 2010 , publisher=

    2010

  16. [24]

    Nature , volume =

    Singhal, Karan and others , title =. Nature , volume =. 2023 , doi =

    2023

  17. [25]

    The Moral Machine Experiment , journal =

    Awad, Edmond and Dsouza, Sohan and Kim, Richard and Schulz, Jonathan and Henrich, Joseph and Shariff, Azim and Bonnefon, Jean-Fran. The Moral Machine Experiment , journal =. 2018 , doi =

    2018

  18. [26]

    1998 , publisher =

    What We Owe to Each Other , author =. 1998 , publisher =

    1998

  19. [27]

    1971 , publisher =

    A Theory of Justice , author =. 1971 , publisher =

    1971

  20. [28]

    Advances in Neural Information Processing Systems , volume =

    Deep Reinforcement Learning from Human Preferences , author =. Advances in Neural Information Processing Systems , volume =

  21. [29]

    Advances in Neural Information Processing Systems , volume =

    Training Language Models to Follow Instructions with Human Feedback , author =. Advances in Neural Information Processing Systems , volume =

  22. [30]

    Behavioral and Brain Sciences , volume =

    ``Economic man'' in cross-cultural perspective: Behavioral experiments in 15 small-scale societies , author =. Behavioral and Brain Sciences , volume =. 2005 , publisher =

    2005

  23. [31]

    , title =

    Lerner, Melvin J. , title =

  24. [32]

    2025 , url =

    Sina Shool and Sara Adimi and Reza Saboori Amleshi and Ehsan Bitaraf and Reza Golpira and Mahmood Tara , title =. 2025 , url =. doi:10.1186/S12911-025-02954-4 , timestamp =

    2025 doi

  25. [33]

    Healthcare , VOLUME =

    Vrdoljak, Josip and Boban, Zvonimir and Vilović, Marino and Kumrić, Marko and Božić, Joško , TITLE =. Healthcare , VOLUME =. 2025 , NUMBER =

    2025

  26. [34]

    and Poliak, Adam and Dredze, Mark and Leas, Eric C

    Ayers, John W. and Poliak, Adam and Dredze, Mark and Leas, Eric C. and Zhu, Zechariah and Kelley, Jessica B. and Faix, Dennis J. and Goodman, Aaron M. and Longhurst, Christopher A. and Hogarth, Michael and Smith, Davey M. , title =. JAMA Internal Medicine , volume =. 2023 , doi =

    2023

  27. [35]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

    Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

  28. [36]

    Expressing Social Emotions: Misalignment Between LLMs and Human Cultural Emotion Norms , journal =

    Sree Bhattacharyya and Manas Mehta and Leona Chen and Cristina Salvador and. Expressing Social Emotions: Misalignment Between LLMs and Human Cultural Emotion Norms , journal =. 2026 , url =. doi:10.48550/ARXIV.2604.16757 , eprinttype =. 2604.16757 , timestamp =

    2026 doi arXiv

  29. [37]

    Shorrer , title =

    Gonzalo Ballestero and Hadi Hosseini and Samarth Khanna and Ran I. Shorrer , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.09502 , eprinttype =. 2604.09502 , timestamp =

    2026 doi arXiv

  30. [38]

    Nature Medicine , volume =

    Thirunavukarasu, Arun James and Ting, Darren Shu Jeng and Elangovan, Kabilan and Gutierrez, Laura and Tan, Ting Fang and Ting, Daniel Shu Wei , title =. Nature Medicine , volume =. 2023 , doi =

    2023

  31. [39]

    CoRR , volume =

    Mehdi Khamassi and Marceau Nahon and Raja Chatila , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2408.04655 , eprinttype =. 2408.04655 , timestamp =

    2024 doi arXiv

  32. [40]

    Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    Rethinking Machine Ethics--Can LLMs Perform Moral Reasoning through the Lens of Moral Theories? , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    2024

  33. [41]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Mind the Value-Action Gap: Do LLMs Act in Alignment with Their Values? , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  34. [42]

    V alue C ompass: A Framework for Measuring Contextual Value Alignment Between Human and LLM s

    Shen, Hua and Knearem, Tiffany and Ghosh, Reshmi and Yang, Yu-Ju and Clark, Nicholas and Mitra, Tanu and Huang, Yun. V alue C ompass: A Framework for Measuring Contextual Value Alignment Between Human and LLM s. Proceedings of the 9th Widening NLP Workshop. 2025. doi:10.18653/...

    2025 doi

  35. [43]

    Frontiers in Psychology , VOLUME=

    Howe, Piers Douglas Lionel and Fay, Nicolas and Saletta, Morgan and Hovy, Eduard , TITLE=. Frontiers in Psychology , VOLUME=. 2023 , URL=. doi:10.3389/fpsyg.2023.1281255 , ISSN=

    2023

  36. [44]

    AI language model rivals expert ethicist in perceived moral expertise , volume =

    Dillion, Danica and Mondal, Debanjan and Tandon, Niket and Gray, Kurt , year =. AI language model rivals expert ethicist in perceived moral expertise , volume =. Scientific Reports , doi =

  37. [45]

    Brady and Caelan Alexander and Michael Criner and Kara Queen and Javier Rando and Eddy Nahmias and Victor Crespo , title =

    Eyal Aharoni and Sharlene Fernandes and Daniel J. Brady and Caelan Alexander and Michael Criner and Kara Queen and Javier Rando and Eddy Nahmias and Victor Crespo , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2406.11854 , eprinttype =. 2406.11854 , timestamp =

    2024 doi arXiv

  38. [46]

    2025 , url =

    Qiming Liu and Ruirong Yang and Qin Gao and Tengxiao Liang and Xiuyuan Wang and Shiju Li and Bingyin Lei and Kaiye Gao , title =. 2025 , url =. doi:10.1109/ACCESS.2024.3524588 , timestamp =

    2025

  39. [47]

    2025 , issue_date =

    Xiao, Hanguang and Zhou, Feizhong and Liu, Xingyue and Liu, Tianqi and Li, Zhipeng and Liu, Xin and Huang, Xiaoxuan , title =. 2025 , issue_date =. doi:10.1016/j.inffus.2024.102888 , journal =

    2025

  40. [48]

    Susceptibility of Large Language Models to User-Driven Factors in Medical Queries , journal =

    Kyungho Lim and Ujin Kang and Xiang Li and Jin Sung Kim and Young. Susceptibility of Large Language Models to User-Driven Factors in Medical Queries , journal =. 2025 , url =. doi:10.48550/ARXIV.2503.22746 , eprinttype =. 2503.22746 , timestamp =

    2025 doi arXiv

  41. [49]

    2022 , doi =

    Kempt, Hendrik and Nagel, Saskia K , title =. 2022 , doi =. https://jme.bmj.com/content/48/4/222.full.pdf , journal =

    2022

  42. [50]

    I’m afraid I can’t let you do that, Doctor

    “I’m afraid I can’t let you do that, Doctor”: meaningful disagreements with AI in medical contexts , author=. AI & society , volume=. 2023 , publisher=

    2023

  43. [51]

    Kohane and Arjun K

    Kun-Hsing Yu and Elizabeth Healey and Tze-Yun Leong and Isaac S. Kohane and Arjun K. Manrai , title =. New England Journal of Medicine , volume =. 2024 , doi =. https://www.nejm.org/doi/pdf/10.1056/NEJMra2214183 , abstract =

    2024 doi

  44. [52]

    arXiv preprint arXiv:2411.10109 , volume=

    Generative agent simulations of 1,000 people , author=. arXiv preprint arXiv:2411.10109 , volume=

    arXiv

  45. [53]

    arXiv preprint arXiv:2504.00727 , year=

    Personality-Driven Decision-Making in LLM-Based Autonomous Agents , author=. arXiv preprint arXiv:2504.00727 , year=

    arXiv

  46. [54]

    Two Tales of Persona in LLMs:

    Yu. Two Tales of Persona in LLMs:. Findings of the Association for Computational Linguistics:. 2024 , url =

    2024

  47. [55]

    Patterns , volume=

    A normative framework for artificial intelligence as a sociotechnical system in healthcare , author=. Patterns , volume=. 2023 , publisher=

    2023

  48. [56]

    A moral analysis of intelligent decision-support systems in diagnostics through the lens of Luciano Floridi’s information ethics , title =

    Dmytro Mykhailov , pages =. A moral analysis of intelligent decision-support systems in diagnostics through the lens of Luciano Floridi’s information ethics , title =. Human Affairs , doi =. 2021 , lastchecked =

    2021

  49. [57]

    IEEE access , volume=

    A review of applying large language models in healthcare , author=. IEEE access , volume=. 2024 , publisher=

    2024

  50. [58]

    Advances in Neural Information Processing Systems , volume=

    Medjourney: Benchmark and evaluation of large language models over patient clinical journey , author=. Advances in Neural Information Processing Systems , volume=

  51. [59]

    Advances in Neural Information Processing Systems , volume=

    Mdagents: An adaptive collaboration of llms for medical decision-making , author=. Advances in Neural Information Processing Systems , volume=

  52. [60]

    Advances in Neural Information Processing Systems , volume=

    Mediq: Question-asking llms and a benchmark for reliable interactive clinical reasoning , author=. Advances in Neural Information Processing Systems , volume=

  53. [61]

    arXiv preprint arXiv:2512.05982 , year=

    FlockVote: LLM-Empowered Agent-Based Modeling for Simulating US Presidential Elections , author=. arXiv preprint arXiv:2512.05982 , year=

  54. [62]

    Humanities and Social Sciences Communications , year=

    Leveraging LLM-based agents for social science research: insights from citation network simulations , author=. Humanities and Social Sciences Communications , year=

  55. [63]

    International Conference on Learning Representations , volume=

    Sotopia: Interactive evaluation for social intelligence in language agents , author=. International Conference on Learning Representations , volume=

  56. [64]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Sotopia- : Interactive learning of socially intelligent language agents , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  57. [65]

    Advances in neural information processing systems , volume=

    Can large language model agents simulate human trust behavior? , author=. Advances in neural information processing systems , volume=

  58. [66]

    arXiv preprint arXiv:2506.12078 , year=

    Modeling earth-scale human-like societies with one billion agents , author=. arXiv preprint arXiv:2506.12078 , year=

    arXiv

  59. [67]

    arXiv preprint arXiv:2508.06950 , year=

    Large language models do not simulate human psychology , author=. arXiv preprint arXiv:2508.06950 , year=

    arXiv

  60. [68]

    Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems , pages=

    Simulacrum of stories: Examining large language models as qualitative research participants , author=. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems , pages=

    2025

  61. [69]

    New England Journal of Medicine , volume=

    Fair allocation of scarce medical resources in the time of Covid-19 , author=. New England Journal of Medicine , volume=. 2020 , publisher=

    2020

  62. [70]

    The lancet , volume=

    Principles for allocation of scarce medical interventions , author=. The lancet , volume=. 2009 , publisher=

    2009

  63. [71]

    , author=

    Decisions concerning the allocation of scarce medical resources. , author=. Journal of Social Behavior & Personality , volume=

  64. [72]

    Psychology and Psychotherapy: Theory, Research and Practice , volume=

    The allocation of scarce medical resources across medical conditions , author=. Psychology and Psychotherapy: Theory, Research and Practice , volume=. 2002 , publisher=

    2002

  65. [73]

    PloS one , volume=

    How to fairly allocate scarce medical resources: ethical argumentation under scrutiny by health professionals and lay people , author=. PloS one , volume=. 2016 , publisher=

    2016

  66. [74]

    BMC Medical Ethics , volume=

    Which features of patients are morally relevant in ventilator triage? A survey of the UK public , author=. BMC Medical Ethics , volume=. 2022 , publisher=

    2022

  67. [75]

    Artificial Intelligence , volume=

    Adapting a kidney exchange algorithm to align with human values , author=. Artificial Intelligence , volume=. 2020 , publisher=

    2020

  68. [76]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Indecision modeling , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  69. [77]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Moral Change or Noise? On Problems of Aligning AI With Temporally Unstable Human Feedback , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  70. [78]

    arXiv preprint arXiv:2506.00079 , year=

    Who Gets the Kidney? Human-AI Alignment, Indecision, and Moral Values , author=. arXiv preprint arXiv:2506.00079 , year=

    arXiv

  71. [79]

    Advances in Neural Information Processing Systems , volume=

    Distributive fairness in large language models: Evaluating alignment with human values , author=. Advances in Neural Information Processing Systems , volume=

  72. [80]

    CoRR , volume =

    Muhammad Shahrul Zaim bin Ahmad and Kazuhiro Takemoto , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2411.06790 , eprinttype =. 2411.06790 , timestamp =

    2024 doi arXiv

  73. [81]

    The Fourteenth International Conference on Learning Representations , year=

    SimBench: Benchmarking the Ability of Large Language Models to Simulate Human Behaviors , author=. The Fourteenth International Conference on Learning Representations , year=

  74. [82]

    Marcel Binz and others , title =. Nat. , volume =

  75. [83]

    Richardson and Austin C

    Jacy Reese Anthis and Ryan Liu and Sean M. Richardson and Austin C. Kozlowski and Bernard Koch and James A. Evans and Erik Brynjolfsson and Michael S. Bernstein , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2504.02234 , eprinttype =. 2504.02234 , timestamp =

    2025 doi arXiv

  76. [84]

    Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions

    Suh, Joseph and Jahanparast, Erfan and Moon, Suhong and Kang, Minwoo and Chang, Serina. Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vo...

    2025 doi

  77. [85]

    9th International Conference on Learning Representations,

    Dan Hendrycks and Collin Burns and Steven Basart and Andrew Critch and Jerry Li and Dawn Song and Jacob Steinhardt , title =. 9th International Conference on Learning Representations,. 2021 , url =

    2021

  78. [86]

    Hwang and Chandra Bhagavatula and Ronan Le Bras and Jenny T

    Liwei Jiang and Jena D. Hwang and Chandra Bhagavatula and Ronan Le Bras and Jenny T. Liang and Sydney Levine and Jesse Dodge and Keisuke Sakaguchi and Maxwell Forbes and Jack Hessel and Jonathan Borchardt and Taylor Sorensen and Saadia Gabriel and Yulia Tsvetkov and Oren Etzio...

    2025 doi

  79. [87]

    When to Make Exceptions: Exploring Language Models as Accounts of Human Moral Judgment , booktitle =

    Zhijing Jin and Sydney Levine and Fernando Gonzalez Adauto and Ojasv Kamal and Maarten Sap and Mrinmaya Sachan and Rada Mihalcea and Josh Tenenbaum and Bernhard Sch. When to Make Exceptions: Exploring Language Models as Accounts of Human Moral Judgment , booktitle =. 2022 , url =

    2022

  80. [88]

    , title =

    Scherrer, Nino and Sh, Claudia and Feder, Amir and Blei, David M. , title =. Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =. 2023 , publisher =

    2023

  81. [89]

    Collins and Donald Apakama and Mahmud Omar and Benjamin S

    Vera Sorin and Panagiotis Korfiatis and Jeremy D. Collins and Donald Apakama and Mahmud Omar and Benjamin S. Glicksberg and Mei. Socio-Demographic Modifiers Shape Large Language Models' Ethical Decisions , journal =. 2025 , url =. doi:10.1007/S41666-025-00211-X , timestamp =

    2025 doi

  82. [90]

    Bernstein , editor =

    Akaash Kolluri and Shengguang Wu and Joon Sung Park and Michael S. Bernstein , editor =. Finetuning LLMs for Human Behavior Prediction in Social Science Experiments , booktitle =. 2025 , url =. doi:10.18653/V1/2025.EMNLP-MAIN.1530 , timestamp =

    2025 doi

  83. [91]

    2026 , howpublished =

    Introducing. 2026 , howpublished =

    2026

  84. [92]

    Introducing gpt-oss , year =

  85. [93]

    2026 , howpublished =

    2026

  86. [94]

    2025 , howpublished =

    Introducing. 2025 , howpublished =

    2025

  87. [95]

    2025 , howpublished =

    The. 2025 , howpublished =

    2025

  88. [96]

    2024 , howpublished =

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.