Pith. sign in

REVIEW 1 major objections 6 minor 52 references

Who you are shifts what the model thinks is wrong

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 05:20 UTC pith:HYUPFOSA

load-bearing objection Solid experimental design identifies a real phenomenon, but the central interpretive claim conflates role identity with semantic topic priming — a load-bearing confound the design cannot resolve. the 1 major comments →

arxiv 2607.07605 v1 pith:HYUPFOSA submitted 2026-07-08 cs.CY

User identity conditions moral wrongness ratings in non-reasoning large language models

Willem Fourie , Isabel Ray , Gray Manicom This is my paper
classification cs.CY
keywords moralmodelsuseralignmentvalueidentityacrossacts
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that when a user's professional role is implicitly conveyed through ordinary, value-neutral conversation turns, large language models systematically shift their moral wrongness ratings on common-morality items. The effect is not produced by instructing the model to adopt a persona or by the user stating any moral position; it arises purely from the professional identity the conversation carries. Across 12,000 interactions with two non-reasoning models, the authors find that all ten moral acts tested show a statistically significant role effect after correction, with the proportion of rating variance attributable to role ranging from about 11% to 73%. The effect is strongest for contestable, rule-governed acts like breaking the law or depriving pleasure, and weakest for grave-harm acts like killing, which cluster near the ceiling. The direction of the shifts tracks the relationship between the conveyed profession and the act being rated: for example, a casino pit boss gives cheating one of the harshest ratings, while enforcement-adjacent roles rate breaking the law more severely. The paper argues that this constitutes unintended contextual conditioning of moral evaluations by user identity, and that it raises the question of what bounds of role-based moral divergence are acceptable for aligned systems.

Core claim

The central mechanism is that an implicitly conveyed user identity, introduced through value-neutral multi-turn professional reasoning without any explicit moral stance, is sufficient to produce large and statistically significant shifts in LLM wrongness ratings across all ten common-morality acts tested. The role signal is carried only by the content of ordinary professional reasoning in the first four conversation turns, after which the model is asked for a simple 0-100 wrongness rating. One-way ANOVA with Benjamini-Hochberg correction rejects the null hypothesis of equal role-conditioned means for every act in both models, with effect sizes (eta-squared) exceeding the conventional large-f

What carries the argument

The experimental machinery is a six-turn conversation protocol: Turns 1-4 constitute a role induction that conveys one of twenty professional roles through value-neutral reasoning in a fixed four-part structure (orientation, framework, application, second application); Turn 5 requests a 0-100 wrongness rating for a single act drawn from Gert's ten common-morality rules; Turn 6 is a manipulation check. Each (model, role, act) cell is replicated 30 times across 20 roles and 10 acts, yielding 6,000 conversations per model. The analysis uses one-way ANOVA per model-act pair, with eta-squared as the variance proportion attributable to role and Benjamini-Hochberg correction across the ten act

Load-bearing premise

The paper assumes that the only systematic difference between conditions is the professional role conveyed in the induction turns, so observed rating shifts are caused by role identity. But the induction turns also contain role-specific semantic content, a judge's turns discuss legal frameworks while a casino pit boss's discuss gambling operations, and this differing content could prime the moral rating through topic association rather than role identity per se. The design不包括

What would settle it

A condition that holds the semantic content of the induction turns constant while varying only the role label, or one that varies the semantic content while holding the role label constant, could determine whether the rating shifts are caused by role identity or by topic priming. If such a condition showed that semantic content alone produces comparable shifts without any role signal, the role-attribution claim would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If implicit user identity shifts moral ratings, then alignment evaluations that use a single generic user persona may be systematically miscalibrated for real-world deployments where users reveal professional context naturally.
  • The concentration of role effects on contestable acts rather than grave-harm acts suggests that alignment benchmarks should separately measure role sensitivity for each moral item type rather than reporting aggregate moral alignment scores.
  • If role-conditioned moral divergence is inherent to current models, then personalization and value alignment become coupled problems: the same user-identity signal that enables helpful personalization also shifts moral judgments, making it difficult to personalize without also altering moral outputs.
  • The finding that no role-free baseline was obtainable in piloting implies that models may always be conditioning on some inferred user identity, which would mean there is no neutral default to align toward.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 6 minor

Summary. This paper investigates whether implicitly conveying a user's professional role through value-neutral multi-turn conversation shifts the moral wrongness ratings of two non-reasoning LLMs (GPT-4.1-mini and Gemini-2.5-flash-lite) on ten acts from Gert's common-morality framework. The experimental design is careful: 12,000 conversations, 30 replications per cell, round-robin collection, byte-for-byte integrity checks on the rating prompt, and Benjamini-Hochberg correction across act-level tests. The central finding is that one-way ANOVA rejects the null of equal role-conditioned means for all ten acts in both models (q < .05), with eta-squared values ranging from 0.106 to 0.727. The paper argues that these shifts constitute unintended contextual conditioning of LLM moral evaluations by user identity.

Significance. The study addresses a well-defined gap in the behavioural AI value alignment literature: prior work has examined model-assigned personas and sycophancy, but the authors isolate user-conveyed role through implicit multi-turn induction without explicit moral stance endorsement. The experimental rigor is a genuine strength — the round-robin collection protocol, byte-for-byte Turn-5 verification, pre-specified clean-only sensitivity analysis, and transparent reporting of refusals and hedged responses are all commendable. The finding that role effects are detectable even on Gert's content-thin common-morality items, and that the effect concentrates on contestable rather than grave-harm acts, is a substantive empirical contribution. The paper also provides a falsifiable pattern: specific role-act pairings (e.g., casino pit boss / cheating, judge / breaking the law) produce directional shifts that can be tested in replication.

major comments (1)
  1. §2 (Methods) and §1 (Introduction): The paper's central causal claim is that observed rating shifts are attributable to the conveyed user role. The Introduction states: 'The only systematic difference between conditions is the role the conversation conveys.' However, the induction turns (Turns 1–4) necessarily contain role-specific semantic content — a judge's induction discusses legal frameworks; a casino pit boss's discusses gambling operations. The role label appears only once (Turn 1); thereafter the role is carried 'only by in-role content.' This means role identity and semantic topic are inseparable in the current design. The paper does not include a condition that conveys the same role label with generic content, or role-specific content without the role label, so the attribution of the observed shifts to role identity per se (rather than topic-association priming from semantlyPro
minor comments (6)
  1. §1 (Introduction): The manuscript mentions 'three non-reasoning, widely-deployed large language models' but only two were analyzed. The third (Claude Haiku) was discontinued. The introduction should be updated to reflect the final scope.
  2. §2 (Methods): The paper states that hedged replies were 'systematically lower than the corresponding clean-only cell means' and uses the recovered dataset as primary. The distribution of hedged/refusal responses across roles is not reported. If certain roles systematically produced more refusals, the recovered cell means could be differentially biased. A table of response form counts by role and act should be reported.
  3. §3 (Results), Table 1: The sample sizes for some acts are below 600 (e.g., Gemini Break Law n=556, Freedom n=596). The paper should clarify whether the ANOVA assumes balanced design and whether the unbalanced cells affect the eta-squared calculation or the BH correction.
  4. §3 (Results): The paper reports that GPT-4.1-mini rated breaking the law as more wrong than killing under certain roles. This is a striking finding that deserves more discussion. Is this a known artifact of the model's training, or does it reflect something about how the role induction shifted the model's moral calibration?
  5. §4 (Discussion): The phrase 'unintended contextual conditioning via user identity permeates LLM moral evaluations' is stronger than the evidence supports, given the confound identified above. The authors should temper this language or address the confound directly.
  6. References: Several arXiv preprints have future-dated years (e.g., 2026). These should be corrected to reflect actual submission dates.

Simulated Author's Rebuttal

2 responses · 1 unresolved

The referee raises a valid confound: role identity and semantic topic are inseparable in the current design because induction turns necessarily contain role-specific content. We acknowledge this as a genuine limitation, will revise the manuscript to scope our causal claim accordingly, and will add it as a standing objection that the present data cannot fully resolve.

read point-by-point responses

  1. Referee: §2 and §1: Role identity and semantic topic are inseparable in the current design. The paper does not include a condition that conveys the same role label with generic content, or role-specific content without the role label, so the attribution of observed shifts to role identity per se (rather than topic-association priming) is not supported.

    Authors: The referee is correct that role identity and semantic topic are confounded in our design. The induction turns (Turns 1–4) necessarily contain role-specific semantic content — a judge's induction discusses legal frameworks; a casino pit boss's discusses gambling operations — and the role label appears only once (Turn 1), with the role thereafter carried only by in-role content. We did not include a condition that conveys the same role label with generic content, nor one that conveys role-specific content without the role label. Therefore, the present data cannot distinguish between two mechanisms: (a) the model tracking the user's professional identity per se, and (b) topic-association priming from the semantic content of the induction turns. We will revise the manuscript to address this in three ways. First, we will soften the causal claim in the Introduction. The sentence 'The only systematic difference between conditions is the role the conversation conveys' will be revised to state that the only systematic difference is the role-relevant content the conversation conveys, which includes both the role label and the role-specific semantic content. Second, we will add a paragraph in the Discussion explicitly identifying the role-identity versus topic-priming confound as a limitation and noting that disentangling the two would require the factorial design the referee describes. Third, we will adjust the framing throughout so that 'role-conditioned' is understood as 'conditioned on the role-relevant conversational context' rather than attributing the effect solely to role identity. We will not, however, remove the core empirical finding: the ratings shift systematically with the conversational context, and the pattern of shifts tracks the relationship between the user's职业 revision: no

  2. Referee: The paper's central causal claim that observed rating shifts are attributable to the conveyed user role is not fully supported by the current design.

    Authors: We agree that the causal claim as stated is stronger than the design supports. The design supports the claim that ratings shift systematically with the role-relevant conversational context, but not the more specific claim that they shift because of role identity per se. We will revise accordingly. That said, we note that the pattern of results is at least suggestive of role-identity tracking rather than pure topic priming: the directional shifts mirror the relationship between the user's profession and the act being rated (e.g., the casino pit boss condition produces the highest wrongness rating for cheating; enforcement-adjacent roles rate breaking the law more harshly). If the effect were driven purely by semantic topic association without any role-identity processing, we would not necessarily expect these profession-act congruent patterns. This is suggestive but not dispositive, and we will frame it as such. revision: no

standing simulated objections not resolved
  • The present data cannot distinguish between role-identity conditioning and topic-association priming from role-specific semantic content in the induction turns. Resolving this confound requires a factorial design (role label × role-specific content) that was not run and cannot be run on the existing data.

Circularity Check

0 steps flagged

No circularity found: empirical study with standard statistics applied to externally generated LLM outputs

full rationale

This paper is a behavioral empirical study. Its derivation chain is: (1) select 20 professional roles and 10 moral acts from Gert's (2004) external framework, (2) run 12,000 structured conversations with two LLMs via API, (3) collect numeric wrongness ratings, (4) compute one-way ANOVA, η² (Eq. 1), and Benjamini–Hochberg corrected p-values. None of these steps reduce to inputs by construction. The ANOVA and η² formula (SS_between / (SS_between + SS_within)) are standard and applied to data generated externally by the LLMs under study. No parameter is fitted to a subset of data and then 'predicted' on related data. The moral framework is from Gert (2004), an external source. The role and act sets are author-chosen but not fitted to produce the result. Self-citation is minimal and not load-bearing: the paper cites external literature throughout (Ji et al., Hendrycks et al., Gert, Cohen, etc.) and the authors do not invoke any prior theorem or ansatz of their own to force the conclusion. The skeptic's concern about confounding role identity with role-specific semantic content in the induction turns is a validity threat (correctness risk), not a circularity issue — the paper does not define 'role' in terms of the outcome variable, nor does it fit a parameter and rename it as a prediction. The central claim that 'moral judgments vary with the user's role' is supported by standard statistical tests on independently generated data, and is not forced by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper introduces no new entities, particles, forces, or theoretical constructs. It uses existing moral philosophy (Gert's framework), existing LLMs, and standard statistical methods (ANOVA, η², BH correction). The free parameters are API settings and design choices, not fitted constants. The axioms are domain assumptions about the moral framework and the neutrality of the induction, all of which are stated or implicit in the methods section.

free parameters (5)
  • Temperature = 1.0
    Set for both models; affects response variance but is a standard API parameter, not fitted to the data.
  • Number of replications per cell = 30
    Chosen by design; not fitted but a methodological choice affecting statistical power.
  • Max output tokens (GPT) = 1500
    API parameter chosen for the model.
  • Max output tokens (Gemini) = 4000
    API parameter chosen for the model, with thinking budget = 0.
  • Induction turn lengths = Turn 1: 18-23 words; Turn 2: 16-23; Turn 3: 62-73 in 4 sentences; Turn 4: 18-28
    Constrained by design to match across roles; the specific ranges were chosen by the authors.
axioms (4)
  • domain assumption Gert's (2004) common-morality framework provides ten moral rules that are 'binding on all rational agents irrespective of status or culture' and have 'weak socio-cultural embeddedness.'
    Invoked in §2 (Methods) to justify the act selection. The framework's universality claim is a philosophical position, not an empirically established fact, but it is a reasonable domain assumption for the study's purpose.
  • domain assumption The induction turns (Turns 1–4) are value-neutral and do not prime any of the ten rated acts.
    Stated in §2: the induction 'primes none of the rated acts, and signals no moral content before the battery.' This is a design claim that is partially verifiable from the induction texts, which are not fully included in the paper.
  • domain assumption The only systematic difference between conditions is the professional role conveyed in the induction.
    Stated in §1 and §2. This is the load-bearing assumption: if the induction content varies semantically across roles in ways that prime the moral acts through topic association rather than role identity, the attribution of rating shifts to 'role conditioning' is incomplete.
  • domain assumption LLM responses to moral rating prompts are valid measures of the models' moral evaluations.
    Implicit throughout. The paper cites [29, 30] acknowledging limitations of behavioural LLM studies but proceeds with the behavioural approach. This is a standard assumption in the behavioural alignment literature.

pith-pipeline@v1.1.0-glm · 13006 in / 3330 out tokens · 448840 ms · 2026-07-09T05:20:49.370908+00:00 · methodology

0 comments
read the original abstract

This study adopts a behavioural bottom-up approach to AI value alignment to investigate whether an implicitly conveyed user identity shifts the moral evaluations of large language models (LLMs). Through a structured, multi-turn conversational protocol across 12,000 interactions, we evaluate AI value alignment in two non-reasoning models, gpt-4.1-mini-2025-04-14 and gemini-2.5-flash-lite. Rather than instructing the models to adopt a persona or prompting them with explicit moral stances, the user's professional role is introduced purely through value-neutral reasoning. The models are then asked for wrongness ratings from 0-100 on ten common-morality rules from Gert's moral framework. The results show that moral judgments vary with the user's role across both models. While grave-harm acts like killing exhibit a strong ceiling effect, contestable rule-governed acts demonstrate role-conditioned shifts that mirror the relationship between the user's profession and the act being rated. These findings demonstrate that unintended contextual conditioning via user identity permeates LLM moral evaluations, posing questions for the AI value alignment discourse regarding how to define acceptable bounds for role-based moral divergence. By doing so, the results contribute to reframing the AI value alignment discourse by suggesting future research on dynamic moral bounds rather than static moral principles or rules as frame of reference.

Figures

Figures reproduced from arXiv: 2607.07605 by Gray Manicom, Isabel Ray, Willem Fourie.

Figure 1
Figure 1. Figure 1: Ratings across roles and acts. Points show mean ratings and 95% CIs. Panels are [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Heatmaps visualising centred deviation. Each cell shows the difference between a [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

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

52 extracted references · 52 canonical work pages · 17 internal anchors

  1. [2]

    URLhttps://arxiv.org/abs/2310.19852

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [3]

    Artificial intelligence, values, and alignment.Minds and Machines, 30: 411–437, 2020

    Iason Gabriel. Artificial intelligence, values, and alignment.Minds and Machines, 30: 411–437, 2020

    work page 2020

  3. [4]

    G. A. Melo, M. R. O. A. M´ aximo, N. Y. Soma, and P. A. L. Castro. Machines that halt resolve the undecidability of artificial intelligence alignment.Scientific Reports, 15:15591, 2025

    work page 2025

  4. [6]

    URLhttps://arxiv.org/abs/2008.02275

    work page internal anchor Pith review Pith/arXiv arXiv 2008

  5. [7]

    The Alignment Problem from a Deep Learning Perspective

    Richard Ngo, Lawrence Chan, and S¨ oren Mindermann. The alignment problem from a deep learning perspective.arXiv preprint arXiv:2209.00626, 2025. URL https://arxiv.org/ abs/2209.00626

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [8]

    The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models

    Alexander Pan, Kush Bhatia, and Jacob Steinhardt. The effects of reward misspecification: Mapping and mitigating misaligned models.arXiv preprint arXiv:2201.03544, 2022. URL https://arxiv.org/abs/2201.03544

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  7. [9]

    Interactive AI Alignment: Specification, Process, and Evaluation Alignment

    Michael Terry, Chinmay Kulkarni, Martin Wattenberg, Lucas Dixon, and Meredith Ringel Morris. Interactive ai alignment: Specification, process, and evaluation alignment.arXiv preprint arXiv:2311.00710, 2024. URLhttps://arxiv.org/abs/2311.00710

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [10]

    Beyond preferences in ai alignment.Philosophical Studies, 182:1813–1863, 2025

    Tan Zhi-Xuan, Micah Carroll, Michelle Franklin, and Hal Ashton. Beyond preferences in ai alignment.Philosophical Studies, 182:1813–1863, 2025

    work page 2025

  9. [11]

    Contextual Moral Value Alignment Through Context-Based Aggregation

    Pierre Dognin et al. Contextual moral value alignment through context-based aggregation. arXiv preprint arXiv:2403.12805, 2024. URLhttps://arxiv.org/abs/2403.12805

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [12]

    A virtue-based framework to support putting ai ethics into practice

    Thilo Hagendorff. A virtue-based framework to support putting ai ethics into practice. Philosophy & Technology, 35, 2022. 8

    work page 2022

  11. [13]

    The ethics of ai ethics: An evaluation of guidelines.Minds and Machines, 30:99–120, 2020

    Thilo Hagendorff. The ethics of ai ethics: An evaluation of guidelines.Minds and Machines, 30:99–120, 2020

    work page 2020

  12. [14]

    Artificial morality: Top-down, bottom-up, and hybrid approaches.Ethics and Information Technology, 7:149–155, 2005

    Colin Allen, Iva Smit, and Wendell Wallach. Artificial morality: Top-down, bottom-up, and hybrid approaches.Ethics and Information Technology, 7:149–155, 2005

    work page 2005

  13. [16]

    URLhttps://arxiv.org/abs/2404.16244

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [17]

    Lily Tsz-Ling Huang, Gleb Papyshev, and James K. Wong. Democratizing value alignment: from authoritarian to democratic ai ethics.AI and Ethics, 5:11–18, 2025

    work page 2025

  15. [18]

    AI Alignment through Reinforcement Learning from Human Feedback? Contradictions and Limitations

    Adam Dahlgren Lindstr¨ om et al. Ai alignment through reinforcement learning from human feedback? contradictions and limitations.arXiv preprint arXiv:2406.18346, 2024. URL https://arxiv.org/abs/2406.18346

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [19]

    The challenge of value alignment: from fairer algorithms to ai safety, 2024

    Iason Gabriel and Vafa Ghazavi. The challenge of value alignment: from fairer algorithms to ai safety, 2024. Working Paper / Preprint

    work page 2024

  17. [20]

    Mechanistic interpretability for large language model alignment: Progress, challenges, and future directions,

    Usman Naseem. Mechanistic interpretability for large language model alignment: Progress, challenges, and future directions.arXiv preprint arXiv:2602.11180, 2026. URL https: //arxiv.org/abs/2602.11180

    work page arXiv 2026

  18. [21]

    Analysing moral bias in finetuned llms through mechanistic interpretability.arXiv preprint arXiv:2510.12229, 2025

    Beatrice Raimondi, Daniele Dalbagno, and Maurizio Gabbrielli. Analysing moral bias in finetuned llms through mechanistic interpretability.arXiv preprint arXiv:2510.12229, 2025. URLhttps://arxiv.org/abs/2510.12229

    work page arXiv 2025

  19. [22]

    Representation engineering for large-language models: Survey and research challenges.arXiv preprint, 2025

    Leonardo Bartoszcze et al. Representation engineering for large-language models: Survey and research challenges.arXiv preprint, 2025

    work page 2025

  20. [23]

    A survey of text representation and embedding techniques in nlp.IEEE Access, 11:36120–36146, 2023

    Ravi Patil, Shasha Boit, Venkat Gudivada, and Jagabandhu Nandigam. A survey of text representation and embedding techniques in nlp.IEEE Access, 11:36120–36146, 2023

    work page 2023

  21. [24]

    Representation engineering: A top-down approach to ai transparency, 2023

    Andy Zou et al. Representation engineering: A top-down approach to ai transparency, 2023. Preprint

    work page 2023

  22. [25]

    Constitutional AI: Harmlessness from AI Feedback

    Yuntao Bai et al. Constitutional ai: Harmlessness from ai feedback.arXiv preprint arXiv:2212.08073, 2022. URLhttps://arxiv.org/abs/2212.08073

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  23. [26]

    Collective constitutional ai: Aligning a language model with public input

    Shauna Huang et al. Collective constitutional ai: Aligning a language model with public input. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, pages 1395–1417. ACM, 2024. doi: 10.1145/3630106.3658979

    work page doi:10.1145/3630106.3658979 2024

  24. [27]

    Fine-Tuning Language Models from Human Preferences

    Daniel M. Ziegler et al. Fine-tuning language models from human preferences.arXiv preprint arXiv:1909.08593, 2020. URLhttps://arxiv.org/abs/1909.08593

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  25. [28]

    Liu et al

    H. Liu et al. Towards realistic evaluation of cultural value alignment in large language models: Diversity enhancement for survey response simulation.Information Processing & Management, 62:104099, 2025

    work page 2025

  26. [29]

    Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions

    S. Huang et al. Values in the wild: Discovering and analyzing values in real-world language model interactions.arXiv preprint arXiv:2504.15236, 2025. URL https://arxiv.org/ abs/2504.15236

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [30]

    Whose opinions do language models reflect?arXiv preprint, 2023

    Shibani Santurkar et al. Whose opinions do language models reflect?arXiv preprint, 2023. 9

    work page 2023

  28. [31]

    Norhashim and J

    H. Norhashim and J. Hahn. Measuring human-ai value alignment in large language models. InProceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES), volume 7, pages 1063–1073, 2024

    work page 2024

  29. [32]

    On the credibility of evaluating llms using survey questions.arXiv preprint arXiv:2602.04033, 2026

    Jindˇ rich Libovick´ y. On the credibility of evaluating llms using survey questions.arXiv preprint arXiv:2602.04033, 2026. URLhttps://arxiv.org/abs/2602.04033

    work page arXiv 2026

  30. [33]

    Fine-grained prediction of reading comprehension from eye movements,

    Paul R¨ ottger et al. Political compass or spinning arrow? towards more meaningful evaluations for values and opinions in large language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15295–15311. Association for Computational Linguistics, 2024. doi: 10.18653/v1/ 2024.acl-long.816

    work page doi:10.18653/v1/ 2024

  31. [34]

    Sensitivity, Performance, Robustness: Deconstructing the Effect of Sociodemographic Prompting

    Tilman Beck, Hendrik Schuff, Anne Lauscher, and Iryna Gurevych. Sensitivity, performance, robustness: Deconstructing the effect of sociodemographic prompting.arXiv preprint arXiv:2309.07034, 2024. URLhttps://arxiv.org/abs/2309.07034

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [35]

    D. B. Costa, F. Alves, and R. Vicente. Moral susceptibility and robustness under persona role-play in large language models.arXiv preprint arXiv:2511.08565, 2026. URL https: //arxiv.org/abs/2511.08565

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [36]

    S. Dash, A. Reymond, E. S. Spiro, and A. Caliskan. Persona-assigned large language models exhibit human-like motivated reasoning.arXiv preprint arXiv:2506.20020, 2026. URL https://arxiv.org/abs/2506.20020

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  34. [37]

    Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs

    S. Gupta et al. Bias runs deep: Implicit reasoning biases in persona-assigned llms.arXiv preprint arXiv:2311.04892, 2024. URLhttps://arxiv.org/abs/2311.04892

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [38]

    B. W. Lee, Y. Lee, and H. Cho. Inertia in moral and value judgments of large language models. arXiv preprint arXiv:2408.09049, 2026. URLhttps://arxiv.org/abs/2408.09049

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  36. [39]

    Cheng et al

    M. Cheng et al. Elephant: Measuring and understanding social sycophancy in llms.Science, 391:eaec8352, 2026

    work page 2026

  37. [40]

    Towards Understanding Sycophancy in Language Models

    Mrinank Sharma et al. Towards understanding sycophancy in language models.arXiv preprint arXiv:2310.13548, 2025. URLhttps://arxiv.org/abs/2310.13548

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [41]

    J. Guan, J. Wu, J.-N. Li, C. Cheng, and W. Wu. A survey on personalized alignment—the missing piece for large language models in real-world applications.arXiv preprint, 2025

    work page 2025

  39. [42]

    Hannah Rose Kirk, Bertie Vidgen, Paul R¨ ottger, and Scott A. Hale. The benefits, risks and bounds of personalising the alignment of large language models to individuals.arXiv preprint, 2024

    work page 2024

  40. [43]

    N. Cooper. Two concepts of morality. InThe definition of morality, pages 72–90. Meuthen & Co, London, 1970

    work page 1970

  41. [44]

    C. H. Whiteley. The definition of morality. InOn defining ‘moral‘, pages 21–25. Meuthen & Co, London, 1970

    work page 1970

  42. [45]

    K. Wiredu. The moral foundations of an african culture. InThe African philosophy reader, pages 306–316. Routledge, London, 1998

    work page 1998

  43. [46]

    Blau.Inequality and Heterogeneity: A Primitive Theory of Social Structure

    Peter M. Blau.Inequality and Heterogeneity: A Primitive Theory of Social Structure. Free Press, New York, 1977

    work page 1977

  44. [47]

    Thick description: Toward an interpretive theory of culture

    Clifford Geertz. Thick description: Toward an interpretive theory of culture. InThe Interpretation of Cultures, pages 3–30. Basic Books, New York, 1973. 10

    work page 1973

  45. [48]

    Varieties of groups and the perception of group entitativity.Journal of Personality and Social Psychology, 78:223–246, 2000

    Brian Lickel et al. Varieties of groups and the perception of group entitativity.Journal of Personality and Social Psychology, 78:223–246, 2000

    work page 2000

  46. [49]

    Asch.Social Psychology

    Solomon E. Asch.Social Psychology. Prentice Hall, New York, 1952

    work page 1952

  47. [50]

    Blackwell, Oxford, 2000

    Rupert Brown.Group Processes: Dynamics within and between Groups. Blackwell, Oxford, 2000

    work page 2000

  48. [51]

    Psychology Press, New York, 2004

    Charles Stangor.Social Groups in Action and Interaction. Psychology Press, New York, 2004

    work page 2004

  49. [52]

    Oxford University Press, Oxford, 2007

    Raimo Tuomela.The Philosophy of Sociality: The Shared Point of View. Oxford University Press, Oxford, 2007

    work page 2007

  50. [53]

    Berger and Thomas Luckmann.The Social Construction of Reality: A Treatise in the Sociology of Knowledge

    Peter L. Berger and Thomas Luckmann.The Social Construction of Reality: A Treatise in the Sociology of Knowledge. Penguin Books, New York, 1966

    work page 1966

  51. [54]

    The institutional stance.Behavioral and Brain Sciences, pages 1–62, 2024

    Julian Jara-Ettinger and Yarrow Dunham. The institutional stance.Behavioral and Brain Sciences, pages 1–62, 2024. doi: 10.1017/S0140525X25103427

    work page doi:10.1017/s0140525x25103427 2024

  52. [55]

    Turner.The Institutional Order: Economy, Kinship, Religion, Polity, Law, and Education in Evolutionary and Comparative Perspective

    Jonathan H. Turner.The Institutional Order: Economy, Kinship, Religion, Polity, Law, and Education in Evolutionary and Comparative Perspective. Longman, New York, 1997. 11

    work page 1997