acceptodds
Under review as a conference paper at ICLR 2027

Moral Mazes in the Era of Large Language Models

Abstract

Navigating complex social situations, from giving critical feedback without hurting morale to rejecting requests without alienating teammates, is an important part of corporate life. Although large language models (LLMs) are entering the workplace, it is unclear how they will navigate or even reshape these norms. To investigate this question, we created HR Simulator, a game where users play as an HR officer and write emails to tackle challenging workplace scenarios, evaluated with GPT-4o as a judge based on scenario-specific rubrics. We analyze over 600 human and LLM emails and find systematic differences in writing style: LLM emails are more formal and empathetic, and human emails are more varied. Across four LLM judges, human emails receive lower average pass rates than LLM-only emails (8-44% versus 27-68%). Rewriting human drafts with an LLM helps mainly when that LLM is a weak writer on its own: rewrites by Claude 3.5 Haiku and Gemini 2.5 Flash beat those models' own emails in five of eight judge comparisons, compared with three of twelve for stronger writers. On the evaluation side, across 23 judge models, judges show more agreement with one another on emails' outcomes as their capability increases. A deeper analysis of 10 judge models' email rankings shows that specificity and empathy are broadly rewarded, and more capable judges weight them less. Scenario-specific tact has a smaller effect that appears to grow with judge capability. Together, our results suggest that as LLMs increasingly become readers of professional emails, human writing may fail to meet their standards, and human-LLM co-writing may better match their preferences.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
communication, email, large language models, evaluation, llm-as-a-judge, game

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3REVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HR Simulator, a game consisting of five workplace communication scenarios, to study how humans and LLMs write and evaluate professional emails. The authors compare human-written, LLM-written, and LLM-rewritten emails and analyze their tone and evaluation outcomes. They further examine agreement across 23 LLM judges and investigate how communication preferences relate to model capability, using descriptors such as empathy, specificity, and scenario-specific tact. The main findings are that LLMs generally produce more formal and empathetic emails, LLM judges tend to favor LLM-written emails, and higher-capability judges show greater agreement and potentially different preferences regarding tact.

Critical strengths
  1. A relevant research question.

Examining LLMs as both writers and evaluators of workplace communication is worthwhile, particularly as AI-mediated communication becomes more common. The paper draws attention to potentially important differences between human and model preferences.

  1. Extensive comparisons across models and evaluation conditions.

The authors examine multiple writing conditions, 23 LLM judges for agreement analysis, and 10 judges for detailed preference analysis. The inclusion of human annotations and additional controls is valuable.

Critical weaknesses
  1. The paper's novel scientific contribution and broader takeaway remain unclear.

The main observations—that LLMs tend to produce more formal and empathetic text, that LLM judges exhibit systematic stylistic preferences, and that their evaluations differ from human judgments—largely overlap with established findings in LLM generation and evaluation research. Although the paper demonstrates these patterns in workplace email scenarios, it is unclear what substantially new knowledge this setting provides.

The potentially more distinctive finding concerns capability-dependent judge preferences, but its implications are not sufficiently developed or validated. The paper raises the possibility that LLMs may reshape workplace communication norms, yet it does not investigate actual changes in communication behavior or demonstrate consequences of the observed preference differences.

The methodological contribution is also limited. HR Simulator consists of five manually designed scenarios using prompted LLMs to simulate recipients and evaluate outcomes. While this is a reasonable experimental setup, the paper does not establish why it constitutes a reusable research instrument rather than a task-specific collection of simulated scenarios. There is little evidence that the framework produces insights that generalize beyond its predefined situations and evaluation criteria.

Overall, despite the substantial experimental effort, I found it difficult to identify a sufficiently novel and consequential takeaway for the broader ICLR community.

  1. The evidence supporting the most distinctive findings is not sufficiently convincing.

The claim that higher-capability judges exhibit increasingly shared communication preferences relies on relatively modest correlations between judge agreement and model capability (r = .34 for parameter count and r = .39 for Arena Elo). Judge pairs are not independent because individual models appear in multiple pairs, and the analysis does not adequately establish whether the trend generalizes beyond the particular models tested.

The analysis of tact raises further concerns. The tact criteria were partly derived from the judges' own rationales, making the construct insufficiently independent of the preferences it is intended to explain. Although the regression reports a significant capability interaction, the analysis includes only 10 judge models, and clustering by email pair does not fully address dependence at the judge level. The judge-level correlation between tact preference and capability is not statistically conclusive (r = .54, p = .10), and the parameter-count sensitivity analysis does not reproduce the interaction.

Moreover, adding tact improves held-out prediction accuracy only from 78.6% to 79.1%. The authors acknowledge several of these limitations, but the current evidence still supports an exploratory observation rather than a robust finding about how communication preferences change with model capability.

These concerns matter because the capability-dependent analysis appears to be the paper's strongest candidate for a novel contribution. Without more convincing evidence, the remaining results are primarily descriptive observations of already familiar LLM behaviors.

Questions for authors
  1. What is the central novel finding of this work beyond previously documented differences in human and LLM writing and LLM-as-judge preferences? How does HR Simulator enable scientific insights that could not be obtained using existing evaluation setups?

  2. What evidence supports the generalizability and reusability of HR Simulator beyond its five predefined scenarios? Can the authors demonstrate that the observed preferences persist under different scenario designs or evaluation criteria?

  3. Do the capability-dependent convergence and tact findings remain robust under judge-level uncertainty estimation, leave-one-judge-out analyses, or independently defined tact annotations? These results would substantially affect my assessment of the paper's main empirical contribution.

2 comments

Thank you for your comments. Please see our responses below.

Re: Novelty of contributions and broader takeaway.

The main observations [...] largely overlap with established findings in LLM generation and evaluation research.

We agree that LLMs’ stylistic tendencies and self-preference have been documented before. Our contribution shows that they persist in high-stakes workplace communication, a setting where LLMs increasingly read and reply to email on users’ behalf. Here, these tendencies are not only evaluation artifacts but can shape which messages get rewarded.

The methodological contribution is also limited.

HR Simulator’s novel contribution is using a text-based game to study open-ended communication, where success depends on how the reader judges the message. Humans and LLMs write for the same scenarios under the same incentives, which makes their communication comparable. We designed a few scenarios carefully by hand rather than generating many noisy ones. Each scenario targets a general communication skill (declining a request, resolving conflict, reversing a commitment, seeking information, and persuasion) rather than situation-specific knowledge.

Can the authors demonstrate that the observed preferences persist under different scenario designs or evaluation criteria?

The judge rationales show the same pattern across scenarios: weaker judges reward concrete detail and warmth, while stronger judges attend more to strategic choices. For example, in scenario 1, weaker judges reward mentioning the recipient’s higher salary as a concrete appeal, while stronger judges see it as transactional (Table 2). In scenario 2, weaker judges reward prescribing specific creative decisions, while stronger judges warn that this may create resentment (Tables 11–14).

its implications are not sufficiently developed or validated.

Observing how norms change as LLMs are adopted would take time beyond the scope of a single study. However, people already use LLMs to improve their emails. The open question is whether the quality metric is defined by human readers or by LLMs. Our results show that these definitions can differ, and that LLMs increasingly converge on their own definition as they become more capable.

Re: Robustness of capability-dependent findings.

The tact criteria were partly derived from the judges’ own rationales, making the construct insufficiently independent of the preferences it is intended to explain.

Our main preference findings concern specificity and empathy rather than tact. Both have the largest associations with preference, with odds ratios above 1.5 for every judge, and both are weighted less by more capable judges (specificity: 2.51 to 1.55, empathy: 2.78 to 1.88). Specificity is defined independently of judge rationales. Prior work shows LLM judges favor concrete detail (Park et al., 2024) and we show that this preference weakens with capability.

Tact is a post-hoc characterization of how weaker and stronger judges disagree, used to test which effect remains beyond specificity. We therefore derived it from judge rationales by design and label it exploratory in the paper. A small gain is what we would expect if tact accounts for only the remaining differences after specificity and empathy.

Do the capability-dependent convergence and tact findings remain robust under judge-level uncertainty estimation, leave-one-judge-out analyses, or independently defined tact annotations?

Specificity serves as an independently defined check and shows the same capability trend as tact. Tact scores are also consistent across annotators: Gemini 3 Flash, Grok 4 Fast, and GPT-5.2 mostly correlate above 0.7 (Appendix B). Excluding Gemini 3 Flash, which supplied the annotations, as a judge leaves the tact interaction nearly unchanged (0.145 vs. 0.148).

References

Junsoo Park, Seungyeon Jwa, Meiying Ren, Daeyoung Kim, and Sanghyuk Choi. OffsetBias: Leveraging Debiased Data for Tuning Evaluators. Findings of the Association for Computational Linguistics: EMNLP 2024. https://arxiv.org/abs/2407.06551

Thank you for the detailed response. I appreciate the clarification that specificity and empathy are the main preference findings, while tact is intended as an exploratory analysis. The examples of how weaker and stronger judges evaluate the same communication choices differently are interesting, and the additional robustness checks are helpful. I also agree that directly observing long-term changes in workplace communication norms is beyond the scope of this study.

However, my main concern remains the novelty and broader scientific contribution of the findings. While HR Simulator provides a useful setting for studying LLM-mediated communication, I am still not fully convinced that the results offer sufficiently new insights beyond previously documented LLM writing tendencies and evaluation biases. The capability-dependent differences are potentially interesting, but their broader significance remains unclear, particularly given that the evidence comes from five predefined scenarios and their evaluation criteria.

I recognize the substantial experimental effort and the value of examining this problem in a workplace context. Nevertheless, the response has not sufficiently changed my assessment of the paper's scientific contribution or broader implications. I therefore maintain my original score.

Reviewer_8n2pOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The paper introduces HR Simulator, a game setting with five workplace scenarios for studying how humans and LLMs write emails as an HR officer to tackle challenging workplace scenarios.
  • The paper conducts experiments to compare human-written, LLM-written, and LLM-revised emails. The authors observe that LLMs produce more formal and empathetic messages that are generally preferred by model judges; human drafts mainly help weaker writing models.
  • Deeper analysis across 10 judge models' rankings found that stronger judges agree more with one another and tending to place greater weight on situational tact; weaker judges place more weights on specificity and empathy.
Critical strengths
  • The experiment settings are clear and well-structured. Experiments holistically compares human writing, LLM writing, and LLM revisions, then examines preferences across multiple model judges. This helps reveal both how LLMs change communication and what their evaluators reward.
Critical weaknesses
  • As the workplace emails mostly target human recipients, I feel like relying on LLM judgments rather than real recipients’ responses is tricky under the task setup in this paper. The small human evaluation provides some validation, but it remains unclear whether higher-scoring emails actually improve understanding, trust, or workplace outcomes.
  • The practical implications of the findings remain unclear, and I feel like the recommended "human-LLM co-writing" as a solution is already adopted broadly in real world settings. Showing which emails LLM judges prefer provides limited guidance for workplace usage or model training. It would be more interesting to evaluate whether these writing differences affect employees’ subsequent decisions, cooperation, or task performance (e.g. add an additional turn for the recipient to make a decision / complete a task / ... to see how they differ).
Questions for authors

Please see Weaknesses.

2 comments

Thank you for your comments. Please see below for our responses to them.

Re: LLM judgments vs. real recipients.

it remains unclear whether higher-scoring emails actually improve understanding, trust, or workplace outcomes.

We would like to clarify that our goal is not to show that higher-scoring emails improve outcomes with human recipients. We study the setting where LLMs read, judge and respond to workplace email on behalf of users (e.g., Gemini in Gmail, personal assistants). There, the model’s reaction comes before a message reaches a person, so LLM judgment is itself the object of study.

Our human study looks at the differences between LLM and human judgment. GPT-4o disagrees with the human majority on 30% of emails, and it is mostly stricter. It fails 9 emails that humans pass and passes only 3 that humans fail (Table 1). Models that mediate correspondence may therefore reward different emails than people would. Measuring downstream effects on real recipients is a natural next step, and we note it as future work.

Re: Practical implications.

I feel like the recommended “human-LLM co-writing” as a solution is already adopted broadly in real world settings.

We agree that co-writing is already common. Our results suggest it is a surface-level fix: LLM rewriting raises human emails’ pass rates under LLM judges mainly when the rewrite model is a weak writer on its own (Figure 4). It does not address the underlying issue that LLM judgment of communication can differ systematically from human judgment. The main implication is that users should be careful when letting LLMs read and reply to email on their behalf.

Showing which emails LLM judges prefer provides limited guidance for workplace usage or model training.

For model training, our results suggest caution when LLM judges are used as reward signals for communication tasks. Their preferences converge as capability increases and diverge from human judgment, so training on them may further narrow the range of communication styles.

We hope these clarifications address the reviewer’s concerns about the paper’s framing and implications, and we would be grateful if the reviewer would consider them in their assessment.

Thank you for the clarification. I believe my original concerns remain on limited contribution to actually improving real-world applications in HR workflows, but rather opens up more questions unanswered in this topic. Therefore, I maintain my score.

Reviewer_kkYVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses an important and timely research question, and HR Simulator offers a potentially useful framework for studying LLM-mediated workplace communication.

Critical strengths

The paper addresses an important and timely research question, and HR Simulator offers a potentially useful framework for studying LLM-mediated workplace communication.

Critical weaknesses
  1. Limited Ecological Validity and Human Evaluation. The framework relies on GPT-4o to simulate recipients, generate scenario outcomes, and evaluate communication effectiveness. Consequently, the reported pass rates primarily reflect model-based judgments rather than actual human reactions. The human evaluation is also limited to five annotators and 40 emails, while the human writers are online participants without screened HR expertise. The authors should provide stronger human-centered validation and clarify the extent to which the findings reflect real-world workplace communication.

  2. Limited Generalizability and Potential Confounding Factors. The study focuses on five hand-designed, English-language scenarios in a fictional technology workplace. Although the authors evaluate 23 judge models, the main writing-source comparisons use only four judges, and the experiments do not establish whether the observed preferences generalize to other professional contexts or communication tasks. Furthermore, the association between judge capability and agreement may be confounded by model family, generation, and post-training differences. The authors should strengthen the generalizability analysis and better distinguish capability-related effects from these confounding factors.

  3. Potential Circularity in the Analysis of Scenario-Specific Tact. The paper argues that more capable judges place greater emphasis on scenario-specific tact. However, the tact criteria are derived partly from judge rationales and are annotated primarily by Gemini 3 Flash, introducing potential circularity into the analysis. The reported improvement in predictive performance from adding tact is also relatively small, with held-out accuracy increasing from 78.6% to 79.1%. The authors should provide a more independent validation of the tact construct, quantify the robustness of its association with judge capability, and clarify whether the evidence supports the strength of the conclusions drawn.

Questions for authors

Follow the weakness.

1 comment

We thank the reviewer for your comments. Please see below for our rebuttal.

1: Ecological validity and human evaluation.

Consequently, the reported pass rates primarily reflect model-based judgments rather than actual human reactions.

We’d like to clarify the paper’s goal: rather than measuring how well emails work on human recipients, we study the setting where LLMs read, judge, and respond to workplace email on behalf of users (e.g., Gemini in Gmail, personal assistants). There, the model’s reaction comes before a message reaches a person, so LLM judgment itself is the object of study.

Our human study aimed to examine the discrepancy between LLM and human judgment and found that models moderately agree with the human majority and are stricter. Hence models that mediate correspondence may reward different emails than people would, which presents a concern about using LLMs to mediate communication.

the human writers are online participants without screened HR expertise.

We did not screen players for HR expertise because the scenarios test delicate communication (declining requests, resolving conflict, reversing commitments) that do not necessarily require HR knowledge and skills. We will make this scope more explicit in the paper.

2. Generalizability and confounding factors.

Although the authors evaluate 23 judge models, the main writing-source comparisons use only four judges

The four judges (GPT-4o, GPT-5.2, Claude Sonnet 4.5, and Grok 4 Fast) span three model providers and a range of capability levels. Our main finding is consistent across them: LLM-only emails receive higher pass rates than human emails in 19 of 20 judge-writer comparisons (Figure 4).

the experiments do not establish whether the observed preferences generalize to other professional contexts or communication tasks.

There is a trade-off between carefully designing a few scenarios and automatically generating many that may be noisy. We chose the former so that each scenario targets a distinct communication skill that recurs across professional contexts.

the association between judge capability and agreement may be confounded by model family, generation, and post-training differences.

We agree and note this in Sections 4 and 6. We do not claim a causal effect of scale. Controlling for post-training is difficult because most frontier developers do not disclose their post-training details. Yet the models that serve as personal assistants are typically these frontier closed-source models. Hence, we choose Elo as a proxy for capability to measure models’ preference trend and note its limitations.

3. Potential circularity of the tact analysis.

the tact criteria are derived partly from judge rationales and are annotated primarily by Gemini 3 Flash, introducing potential circularity into the analysis.

Our main preference findings concern specificity and empathy rather than tact. Both have the largest associations with preference, with odds ratios above 1.5 for every judge, and both are weighted less by more capable judges (specificity: 2.51 to 1.55, empathy: 2.78 to 1.88). Specificity is defined independently of judge rationales. Tact is a post-hoc characterization of how weaker and stronger judges disagree, used to test whether any effect remains beyond specificity. We therefore derived it from judge rationales by design and label it exploratory in the paper.

The reported improvement in predictive performance from adding tact is also relatively small

A small gain is what we would expect if tact accounts for only the remaining differences after specificity and empathy. To check that tact does not depend on a single annotator, we note two results. Excluding Gemini 3 Flash as a judge leaves the tact interaction nearly unchanged (0.145 vs. 0.148). Tact scores from Gemini 3 Flash, Grok 4 Fast and GPT-5.2 mostly correlate above 0.7 (Appendix B).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.