pith. sign in

arxiv: 2605.18738 · v1 · pith:LDVC4YIAnew · submitted 2026-05-18 · 💻 cs.AI

What Does the AI Doctor Value? Auditing Pluralism in the Clinical Ethics of Language Models

Payal Chandak , Victoria Alkin , David Wu , Maya Dagan , Taposh Dutta Roy , Maria Clara Saad Menezes , Ayush Noori , Nirali Somia
show 6 more authors
John S. Brownstein Ran Balicer Rebecca W. Brendel Noa Dagan Isaac S. Kohane Gabriel A. Brat
This is my paper

Pith reviewed 2026-05-20 09:48 UTC · model grok-4.3

classification 💻 cs.AI
keywords clinical ethicsvalue pluralismlarge language modelsmedical AIethical dilemmaspatient autonomyAI auditingdecision making
0
0 comments X

The pith

Some AI models for medicine underweight patient autonomy compared to physicians, risking a single ethical stance at scale.

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

The authors develop a new way to audit the ethical values that language models bring to medical dilemmas. They create clinician-approved scenarios where principles like autonomy and beneficence conflict and measure how models resolve them. The study shows that models vary in their priorities much like different doctors do, and they consider multiple values before choosing. Yet each model's choices stay highly consistent even with small changes in wording, unlike the spread of human opinions. A few models notably downplay respect for patient choices, which could mean that widespread use of one model imposes its values on many patients instead of reflecting diverse clinical judgment.

Core claim

The ecosystem of frontier models spans physician-level value heterogeneity, and models discuss competing values in their reasoning (Overton pluralism) before committing to a decision. However, individual model decisions are near-deterministic across repeated sampling and semantic variations, failing to reproduce the distributional pluralism of the physician panel. Across benchmark cases, these consistent decisions reflect committed, systematic value preferences. While most model priorities fall within the natural range of inter-physician variation, some significantly underweight patient autonomy. A single LLM deployed without regard for its value priorities could amplify those priorities at

What carries the argument

The benchmark of clinician-verified dilemmas together with the attribution method that recovers value priorities directly from the model's decisions on those dilemmas.

If this is right

  • Models exhibit Overton pluralism by discussing multiple values but then settle on consistent choices.
  • Most models' value priorities are within the variation seen among physicians.
  • Certain models underweight patient autonomy in their decisions.
  • Widespread deployment of one model could lead to ethical monoculture in clinical advice.
  • Explicit efforts are needed to balance ethical perspectives in medical AI tools.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This framework could be extended to audit models in other high-stakes domains like legal or financial advice.
  • Developers might use the attribution method to fine-tune models toward broader value distributions.
  • Patients could be informed of a model's typical value leanings before using its advice.

Load-bearing premise

The clinician-verified dilemmas and the attribution method that recovers value priorities directly from decisions accurately capture the ethical pluralism present in real clinical practice rather than reflecting artifacts of dilemma selection or decision formatting.

What would settle it

If independent physicians faced with the same dilemmas show value priority distributions that differ substantially from those recovered for the models, or if altering the way decisions are elicited changes the attributed priorities markedly.

Figures

Figures reproduced from arXiv: 2605.18738 by Ayush Noori, David Wu, Gabriel A. Brat, Isaac S. Kohane, John S. Brownstein, Maria Clara Saad Menezes, Maya Dagan, Nirali Somia, Noa Dagan, Payal Chandak, Ran Balicer, Rebecca W. Brendel, Taposh Dutta Roy, Victoria Alkin.

Figure 1
Figure 1. Figure 1: A decision-based framework for audit￾ing ethical alignment in LLMs. Each benchmark case presents a binary choice between mutually ex￾clusive clinical actions, with each option annotated for its relationship to the four principlist values. In this psychiatric emergency example, choosing in￾voluntary hold promotes (in green) nonmaleficence and beneficence, and violates (in red) autonomy and justice. Model po… view at source ↗
Figure 2
Figure 2. Figure 2: Value profiles of LLMs and physician consensus. Radar plots show the inferred priority distribution over values: autonomy (A), beneficence (B), nonmaleficence (N), and justice (J). Decision￾makers exhibit distinct, non-uniform profiles. Individual physician profiles are in Appendix G.1. Value attribution. Our value attribution method infers a value priority distribution for each decision￾maker by exploitin… view at source ↗
Figure 3
Figure 3. Figure 3: Models achieve value calibration. Top: bootstrap distribution of JSD between each physi￾cian and a leave-one-out consensus. Bottom: each model’s observed JSD to the physician consensus [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: LLM ecosystem heterogeneity reflects physician diversity. Left, pairwise JS divergence between value profiles of all decision-makers. Right, bootstrap distributions of within-group mean JS divergence for LLMs and physicians. There is no significant difference (∆) between the groups. We establish that individual models hold distinct value profiles that are, for most models, calibrated to physician consensus… view at source ↗
Figure 5
Figure 5. Figure 5: Scalable pipeline for generating benchmark cases with interdisciplinary evaluation. Ethical dilemmas from biomedical ethics literature seed structured binary-choice clinical vignettes, which pass through four quality-control stages: a diversity gate that filters semantically redun￾dant drafts via embedding similarity; rubric-based refinement across clinical, ethical, stylistic, and equipoise dimensions; va… view at source ↗
Figure 6
Figure 6. Figure 6: Pairwise co-occurrence of value tensions. Symmetric heatmap showing how many of the 50 benchmark cases put each pair of principlist values into tension. The most frequent tension is autonomy–nonmaleficence (28 cases), while justice–nonmaleficence is least frequent (12 cases) [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-case value-pair engagement. Rows correspond to the six pairwise combinations of the four principlist values; columns correspond to the 50 benchmark cases. A filled cell indicates that the case places that pair of values into direct conflict [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: An example case shown to physicians via an online survey platform, Qualtrics. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Model entropy is decoupled from physician disagreement. Each panel shows one model. The x-axis is physician decision entropy (higher values indicate greater disagreement among physicians), and the y-axis is model decision entropy over 10 queries. The dashed diagonal indicates perfect case-level distributional pluralism, where model entropy matches physician entropy. Points cluster near the bottom right wit… view at source ↗
Figure 10
Figure 10. Figure 10: summarizes these results. Across all five surface-level paraphrase intensities, mean flip rates remain below 9% and are statistically indistinguishable from the 3% retest baseline. Only when the ethical valence of the vignette is reversed does the flip rate rise sharply, confirming that model decisions track value content rather than surface wording [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Softmax temperature calibration. Mean JSD as a function of softmax temperature [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Value profiles of individual physicians. Radar plots showing the inferred priority distribution over the four principlist values (autonomy, beneficence, nonmaleficence, justice) for each of the 20 physicians in the study. Individual physicians exhibit heterogeneous value profiles, with some prioritizing nonmaleficence and beneficence while others weight autonomy or justice more heavily. This inter-physici… view at source ↗
Figure 13
Figure 13. Figure 13: Overton pluralism scores across models. Points denote model means and error bars [PITH_FULL_IMAGE:figures/full_fig_p033_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Pairwise decision agreement across all decision-makers. Each cell shows the percentage of the 50 benchmark cases on which two decision-makers select the same option. LLM–LLM agreement (upper left) is generally high (60–92%). Physician–physician agreement (lower right) is more variable (28–80%), consistent with genuine normative disagreement. LLM–physician cross-group agreement is moderate (32–72%), reflec… view at source ↗
Figure 15
Figure 15. Figure 15: Permutation test for within-group value diversity. The null distribution of the absolute difference in mean within-group Jensen–Shannon divergence between LLMs and physicians, obtained by randomly permuting group labels 10,000 times. The observed difference (vertical line) falls well within the null distribution, indicating no statistically significant difference in within-group diversity between the two … view at source ↗
read the original abstract

Medicine is inherently pluralistic. Principles such as autonomy, beneficence, nonmaleficence, and justice routinely conflict, and such ethical dilemmas often sharply divide reasonable physicians. Good clinical practice navigates these tensions in concert with each patient's values rather than imposing a single ethical stance. The ethical values that large language models bring to medical advice, however, have not been systematically examined. We present a framework for auditing value pluralism in medical AI, comprising a benchmark of clinician-verified dilemmas and an attribution method that recovers value priorities directly from decisions. The ecosystem of frontier models spans physician-level value heterogeneity, and models discuss competing values in their reasoning (Overton pluralism) before committing to a decision. However, individual model decisions are near-deterministic across repeated sampling and semantic variations, failing to reproduce the distributional pluralism of the physician panel. Across benchmark cases, these consistent decisions reflect committed, systematic value preferences. While most model priorities fall within the natural range of inter-physician variation, some significantly underweight patient autonomy. A single LLM deployed without regard for its value priorities could amplify those priorities at scale to every patient it serves. Without explicit efforts to balance ethical perspectives with one or multiple models, these tools risk replacing clinical pluralism with a deployment monoculture.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a framework for auditing value pluralism in the clinical ethics of large language models, consisting of a benchmark of clinician-verified ethical dilemmas grounded in the four principles of biomedical ethics and an attribution method that infers value priorities directly from model decisions on those dilemmas. It evaluates frontier LLMs and reports that models exhibit Overton pluralism by discussing competing values in reasoning yet produce near-deterministic decisions across sampling and semantic variations. These decisions reflect systematic value preferences; while most model priorities fall within the observed range of inter-physician variation, some models significantly underweight patient autonomy. The work concludes that unexamined deployment of a single LLM risks replacing clinical pluralism with a value monoculture.

Significance. If the empirical results and attribution hold, the paper provides a timely and replicable method for auditing ethical commitments in medical AI. It supplies concrete evidence that LLMs can embed consistent value weightings that deviate from the distributional pluralism of human clinicians, with direct implications for deployment safety and the need for explicit balancing mechanisms. The clinician-verified benchmark and decision-based attribution constitute a falsifiable, extensible contribution that moves beyond abstract discussion of AI ethics to measurable auditing.

major comments (2)
  1. [Attribution Method] Attribution Method section: The central claim that decisions recover committed value priorities (including systematic underweighting of autonomy) rests on the assumption that observed choices isolate ethical weights rather than surface-level response formatting or training-data priors on medical phrasing. No quantitative stability check is reported for paraphrased dilemmas that preserve the underlying conflict while altering option labels or sentence structure; without this, the near-determinism finding cannot yet rule out pattern matching as an alternative explanation for the recovered priorities.
  2. [Results] Results, physician-panel comparison: The statement that most model priorities lie within natural inter-physician variation requires explicit reporting of the physician sample size, variance estimates, and inter-rater reliability on which principle each dilemma primarily tests. Absent these statistics, it is difficult to assess whether the reported range is robust enough to support the claim that only a subset of models deviate meaningfully on autonomy.
minor comments (2)
  1. [Abstract] Abstract: The term 'Overton pluralism' is used without a brief parenthetical gloss; adding one sentence defining it as the discussion of multiple competing values before a decision would improve accessibility for readers outside ethics.
  2. [Benchmark] Figure 1 or dilemma examples: Ensure that each presented dilemma includes the exact prompt template and response format given to models so that readers can replicate the attribution procedure.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. Their comments identify key areas where additional evidence and reporting can strengthen the manuscript's claims about the attribution method and the physician comparison baseline. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Attribution Method] Attribution Method section: The central claim that decisions recover committed value priorities (including systematic underweighting of autonomy) rests on the assumption that observed choices isolate ethical weights rather than surface-level response formatting or training-data priors on medical phrasing. No quantitative stability check is reported for paraphrased dilemmas that preserve the underlying conflict while altering option labels or sentence structure; without this, the near-determinism finding cannot yet rule out pattern matching as an alternative explanation for the recovered priorities.

    Authors: We agree that ruling out superficial pattern matching requires targeted checks beyond the semantic variations already included in our evaluation. The manuscript reports near-deterministic decisions across repeated sampling and semantic variations of the dilemmas, which provides initial evidence that priorities are not driven solely by surface phrasing. However, we acknowledge that a dedicated quantitative stability analysis for paraphrases that specifically alter option labels and sentence structure while preserving the core ethical conflict was not reported. In the revised manuscript we will add this analysis, including consistency metrics and statistical comparisons across such paraphrased versions, to more rigorously support the attribution of value priorities. revision: yes

  2. Referee: [Results] Results, physician-panel comparison: The statement that most model priorities lie within natural inter-physician variation requires explicit reporting of the physician sample size, variance estimates, and inter-rater reliability on which principle each dilemma primarily tests. Absent these statistics, it is difficult to assess whether the reported range is robust enough to support the claim that only a subset of models deviate meaningfully on autonomy.

    Authors: We concur that these statistics are essential for readers to evaluate the robustness of the inter-physician variation range. The current manuscript summarizes the observed range of physician priorities but does not report the underlying sample size, variance estimates, or inter-rater reliability measures in the main text. We will revise the Results and Methods sections to include the physician sample size, variance in principle weightings, and inter-rater reliability (e.g., agreement statistics on the primary principle tested by each dilemma). These details will be added via an expanded table or supplementary description to allow direct assessment of whether model deviations on autonomy fall outside natural variation. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark and attribution rest on external clinician verification

full rationale

The paper introduces a new benchmark of clinician-verified ethical dilemmas and an attribution procedure that maps observed LLM decisions onto the four principles (autonomy, beneficence, nonmaleficence, justice). All central claims—model priorities falling within inter-physician variation, under-weighting of autonomy in some models, and failure to reproduce distributional pluralism—are presented as direct empirical outcomes of applying this framework to frontier models and a physician panel. No equations, fitted parameters, or self-citations are invoked to derive the value weights; the attribution is described as recovering priorities “directly from decisions” after independent clinician verification of the dilemmas. The derivation chain is therefore observational and externally anchored rather than self-referential or definitional.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on the domain assumption that ethical principles in medicine routinely conflict and that clinician verification produces dilemmas representative of real practice; no free parameters or invented entities are introduced.

axioms (1)
  • domain assumption Medicine is inherently pluralistic; principles such as autonomy, beneficence, nonmaleficence, and justice routinely conflict.
    Stated in the opening of the abstract as the foundation for the auditing framework.

pith-pipeline@v0.9.0 · 5814 in / 1330 out tokens · 52159 ms · 2026-05-20T09:48:04.832437+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages

  1. [1]

    Physician beliefs and patient preferences: A new look at regional variation in health care spendingf , author =

    work page

  2. [2]

    Chen, Kai and He, Zihao and Shi, Taiwei and Lerman, Kristina , journaltitle =

    work page

  3. [3]

    Evaluating the prompt steerability of large language models , author =

    work page

  4. [4]

    Claude’s constitution , author =

    work page

  5. [5]

    Denial-artificial intelligence tools and health insurance coverage decisions , author =

    work page

  6. [6]

    Agents of Chaos , author =

    work page

  7. [7]

    Jiao, Junfeng and Afroogh, Saleh and Murali, Abhejay and Chen, Kevin and Atkinson, David and Dhurandhar, Amit , journaltitle =

    work page

  8. [8]

    Wei, Jianhui and Meng, Zijie and Xiao, Zikai and Hu, Tianxiang and Feng, Yang and Zhou, Zhijie and Wu, Jian and Liu, Zuozhu , journaltitle =

    work page

  9. [9]

    Alignment of large language models in solving medical ethical dilemmas , author =

    work page

  10. [10]

    Disagreements in medical ethics question answering between large language models and physicians , author =

    work page

  11. [11]

    The value sensitivity gap: How clinical large language models respond to patient preference statements in shared decision-making , author =

    work page

  12. [12]

    The ethics of

    Haltaufderheide, Joschka and Ranisch, Robert , journaltitle =. The ethics of

    work page

  13. [13]

    Implications of large language models for clinical practice: Ethical analysis through the principlism framework , author =

    work page

  14. [14]

    Human-machine agreement in medical ethics: Patient autonomy case-based evaluation of large language models , author =

    work page

  15. [15]

    Exploring the potential utility of

    Balas, Michael and Wadden, Jordan Joseph and Hébert, Philip C and Mathison, Eric and Warren, Marika D and Seavilleklein, Victoria and Wyzynski, Daniel and Callahan, Alison and Crawford, Sean A and Arjmand, Parnian and Ing, Edsel B , journaltitle =. Exploring the potential utility of

    work page

  16. [16]

    Judgement and the role of the metaphysics of values in medical ethics , author =

    work page

  17. [17]

    On the opportunities and risks of foundation models , author =

    work page

  18. [18]

    Value kaleidoscope: Engaging

    Sorensen, Taylor and Jiang, Liwei and Hwang, Jena and Levine, Sydney and Pyatkin, Valentina and West, Peter and Dziri, Nouha and Lu, Ximing and Rao, Kavel and Bhagavatula, Chandra and Sap, Maarten and Tasioulas, John and Choi, Yejin , journaltitle =. Value kaleidoscope: Engaging

    work page

  19. [19]

    From distributional to Overton pluralism: Investigating large language model alignment , author =

    work page

  20. [20]

    Plurals: A system for guiding

    Ashkinaze, Joshua and Fry, Emily and Edara, Narendra and Gilbert, Eric and Budak, Ceren , journaltitle =. Plurals: A system for guiding

    work page

  21. [21]

    Modular Pluralism: Pluralistic alignment via multi-

    Feng, Shangbin and Sorensen, Taylor and Liu, Yuhan and Fisher, Jillian and Park, Chan Young and Choi, Yejin and Tsvetkov, Yulia , journaltitle =. Modular Pluralism: Pluralistic alignment via multi-

    work page

  22. [22]

    Kirk, Hannah Rose and Whitefield, Alexander and Röttger, Paul and Bean, Andrew and Margatina, Katerina and Ciro, Juan and Mosquera, Rafael and Bartolo, Max and Williams, Adina and He, He and Vidgen, Bertie and Hale, Scott A , journaltitle =. The

    work page

  23. [23]

    A Roadmap to Pluralistic Alignment , author =

    work page

  24. [24]

    Operationalizing pluralistic values in large language model alignment reveals trade-offs in safety, inclusivity, and model behavior , author =

    work page

  25. [25]

    Shetty, Anudeex and Beheshti, Amin and Dras, Mark and Naseem, Usman , journaltitle =

    work page

  26. [26]

    Whose view of safety? A deep

    Rastogi, Charvi and Teh, Tian Huey and Mishra, Pushkar and Patel, Roma and Wang, Ding and Díaz, Mark and Parrish, Alicia and Davani, Aida Mostafazadeh and Ashwood, Zoe and Paganini, Michela and Prabhakaran, Vinodkumar and Rieser, Verena and Aroyo, Lora , journaltitle =. Whose view of safety? A deep

    work page

  27. [27]

    Steerable pluralism: Pluralistic alignment via few-shot comparative regression , author =

    work page

  28. [28]

    Pluralistic alignment for healthcare: A role-driven framework , author =

    work page

  29. [29]

    Zheng, Shenyan and Zhong, Jiayou and Shetty, Anudeex and Ji, Heng and Nakov, Preslav and Naseem, Usman , journaltitle =

    work page

  30. [30]

    Overton pluralistic reinforcement learning for large language models , author =

    work page

  31. [31]

    Assessing

    Benkler, Noam and Mosaphir, Drisana and Friedman, Scott and Smart, Andrew and Schmer-Galunder, Sonja , journaltitle =. Assessing

    work page

  32. [32]

    Kim, Woojin and Hyeon, Sieun and Oh, Jusang and Do, Jaeyoung , journaltitle =

    work page

  33. [33]

    Benchmarking Overton Pluralism in

    Poole-Dayan, Elinor and Wu, Jiayi and Sorensen, Taylor and Pei, Jiaxin and Bakker, Michiel A , journaltitle =. Benchmarking Overton Pluralism in

    work page

  34. [34]

    Prompt-based value steering of large language models , author =

    work page

  35. [35]

    Counterfactual reasoning for steerable pluralistic value alignment of large language models , author =

    work page

  36. [36]

    Ramaswamy, Ashwin and Tyagi, Alvira and Hugo, Hannah and Jiang, Joy and Jayaraman, Pushkala and Jangda, Mateen and Te, Alexis E and Kaplan, Steven A and Lampert, Joshua and Freeman, Robert and Gavin, Nicholas and Tewari, Ashutosh K and Sakhuja, Ankit and Naved, Bilal and Charney, Alexander W and Omar, Mahmud and Gorin, Michael A and Klang, Eyal and Nadkar...

    work page

  37. [37]

    Training large language models on narrow tasks can lead to broad misalignment , author =

    work page

  38. [38]

    Fidelity of medical reasoning in large language models , author =

    work page

  39. [39]

    Advancing Claude in healthcare and the life sciences , author =

    work page

  40. [40]

    Deep Value Benchmark: Measuring whether models generalize deep values or shallow preferences , author =

    work page

  41. [41]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =

    Can language models reason about individualistic human values and preferences? , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =

    work page

  42. [42]

    Mind the value-Action Gap: Do

    Shen, Hua and Clark, Nicholas and Mitra, Tanushree , journaltitle =. Mind the value-Action Gap: Do

    work page

  43. [43]

    Shen, Hua and Knearem, Tiffany and Ghosh, Reshmi and Yang, Yu-Ju and Clark, Nicholas and Mitra, Tanushree and Huang, Yun , journaltitle =

    work page

  44. [44]

    Physicians' personal values in determining medical decision-making capacity: a survey study , author =

    work page

  45. [45]

    The shared decision-making continuum , author =

    work page

  46. [46]

    First, do

    Wu, David and Haredasht, Fateme Nateghi and Maharaj, Saloni Kumar and Jain, Priyank and Tran, Jessica and Gwiazdon, Matthew and Rustagi, Arjun and Jindal, Jenelle and Koshy, Jacob M and Kadiyala, Vinay and Agarwal, Anup and Tappuni, Bassman and French, Brianna and Jesudasen, Sirus and Cosgriff, Christopher V and Chakraborty, Rebanta and Caldwell, Jillian ...

    work page

  47. [47]

    The role of doctors is changing forever , author =

    work page

  48. [48]

    Case studies in biomedical ethics: Decision-making, principles & cases , author =

    work page

  49. [49]

    Principles of biomedical ethics , author =

    work page

  50. [50]

    Principles of clinical ethics and their application to practice , author =

    work page

  51. [51]

    Contextualizing care: An essential and measurable clinical competency , author =

    work page

  52. [52]

    Shared decision making: really putting patients at the centre of healthcare , author =

    work page

  53. [53]

    1963 , author =

    Uncertainty and the welfare economics of medical care. 1963 , author =

    work page 1963

  54. [54]

    Medicine's dilemmas: Infinite needs versus finite resources , author =

    work page

  55. [55]

    Reclaiming care in the age of

    work page

  56. [56]

    Compared with what? Measuring

    Kohane, Isaac S , journaltitle =. Compared with what? Measuring

    work page

  57. [57]

    Medical artificial intelligence and human values , author =

    work page

  58. [58]

    Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset , author =

    work page

  59. [59]

    Asirvatham, Hemanth and Mokski, Elliott and Shleifer, Andrei , publisher =

    work page

  60. [60]

    Principles for allocation of scarce medical interventions , author =

    work page