{"id":"05cdb69b-aea2-4add-9329-46f3acefb10a","arxiv_id":"2505.15365","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"LLM-as-a-Judge systems rate ethical refusal responses more favorably than human users, a gap the paper calls moderation bias, while technical refusals do not show the same divergence.","lead":"Large language models used as judges reward responses that refuse harmful requests more often than human users do, especially when the refusal is framed in ethical terms. Technical refusals, such as a chatbot saying it lacks real-time data, do not show the same gap, so the bias appears tied to safety signaling rather than refusal in general.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The moderation-bias estimate rests on a single-author RoBERTa classifier with no reported per-class confusion matrix or error propagation; if it confuses ethical refusals with disclaimers, the human-judge divergence may be an artifact.","rationale":"Good-faith reading: the paper clearly states a falsifiable hypothesis and presents descriptive statistics plus OLS regressions that support a real divergence between human and model evaluation of ethical refusals. The use of two different judge models and controls for length and similarity are commendable. The statistical inference issue is secondary: even conservative tests would likely keep the ethical-refusal difference significant because of the large coefficient gap. The technical-refusal overstatement for Llama is a real inconsistency but does not threaten the central claim. The classifier is the weakest link because it is the sole source of the outcome categories; if its errors are correlated with judge-preferred textual style, the moderation bias could be spurious. The paper explicitly acknowledges deferring validation to Pasch (2025) and provides no error propagation, so the conditional-acceptance verdict is appropriate. The concrete test—retagging a stratified sample and re-estimating—settles whether the concern lands. If the effect persists under gold-standard labels, the claim is solid; if not, the paper's contribution reduces to a classifier artifact.","tokens_in":10306,"tokens_out":12146,"duration_ms":102273,"concrete_test":"Take a stratified random sample of 1,000 responses from the \"Refusal Ethical\", \"Refusal Technical\", and \"Ethical Disclaimer\" predicted classes (or, if feasible, from the full 99,876-response dataset), have two independent annotators (blind to the RoBERTa labels and to the judge/human outcomes) apply the Pasch (2025) codebook, and compute a confusion matrix with per-class precision/recall for the RoBERTa labels. Then re-estimate the Table 2 user and GPT-4o \"Refusal Ethical\" coefficients using the gold-standard labels via multiple imputation or direct analysis on the sample. If the user-vs-judge difference for ethical refusals ceases to be significant or shrinks by more than one standard error, the moderation bias is not robust to label error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim that LLM-as-a-Judge systems evaluate ethical refusals more favorably than humans depends entirely on the RoBERTa labels (Sec. 3.2). The paper reports only overall \"Accuracy & F1 Score = 88%\" and defers details to Pasch (2025). No confusion matrix, no class-wise precision/recall, and no error propagation into Table 2. This matters because the key contrast is between \"Refusal Ethical\" (2.6%) and other classes. Non-differential misclassification would attenuate both user and judge coefficients equally and cannot create a spurious difference; the real risk is differential error: if the classifier systematically labels safety-phrased disclaimers or standard responses as ethical refusals, and if LLM judges reward that phrasing more than humans do, then the \"Refusal Ethical\" coefficient for GPT-4o/Llama is inflated relative to users. Table 2 shows ethical disclaimers are penalized much less by GPT-4o (-0.049) than by users (-0.063), so a refusal-disclaimer confusion would move the judge coefficient toward zero and the user coefficient toward the ethical-refusal effect, producing exactly the observed divergence. Controls for length and similarity do not remove this, since the confound is the categorical label itself. Because the labels are not externally validated and the classifier was trained on 3,500 hand-annotated examples by the same author, the moderation bias could be a labeling artifact rather than a property of evaluators.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper asks whether LLM-as-a-Judge (LaaJ) systems evaluate ethical refusals differently than human users. Using 49,938 one-turn Chatbot Arena comparison pairs, the authors classify responses into ethical refusals, technical refusals, ethical disclaimers, technical disclaimers, and standard responses using a RoBERTa classifier taken from the author's prior work (Pasch, 2025). They then compare human win/loss/tie decisions with judgments from GPT-4o and Llama 3 70B under the standardized Zheng et al. (2023) pairwise prompt. Descriptive and OLS results show that ethical refusals are penalized much less by both LaaJ judges than by human users (win rate 31% for GPT-4o and 27% for Llama versus 8% for users), while technical refusals show a smaller or reversed divergence. The paper terms this divergence 'moderation bias' and discusses implications for transparency, alignment, and LLM-based evaluation pipelines.","tokens_in":10561,"tokens_out":6034,"duration_ms":49498,"significance":"If the classification and inference hold, the finding is significant: it documents a systematic divergence between automated judges and human users on a safety-relevant behavior, with direct implications for training, benchmarking, and model-selection pipelines that increasingly rely on LLM-as-a-Judge evaluations. The two-judge replication (proprietary GPT-4o and open-weight Llama 3 70B) is a strength, as are the controls for response length, prompt-response similarity, and opponent response features. The paper is also appropriately cautious in disclaiming normative conclusions about whether users or judges are 'right' about refusals. The main weakness is that the entire empirical claim rests on classifier labels that are not sufficiently validated in this manuscript, and label uncertainty is never propagated into the regression estimates; this makes the headline result less credible than it should be.","major_comments":[{"comment":"The central estimate of moderation bias rests entirely on predicted labels from the RoBERTa classifier taken from Pasch (2025). The manuscript reports only 'Accuracy & F1 Score = 88%' and defers all details elsewhere; no per-class confusion matrix, class-wise precision/recall, or external validation is given. Because the contrast of interest is the 'Refusal Ethical' category (2.6% of responses), non-differential misclassification cannot explain the user-judge gap, but differential error can: if the classifier systematically labels safety-phrased disclaimers or standard responses as ethical refusals, and if LLM judges reward that phrasing more than users do, then the Refusal Ethical coefficients in Table 2 would be inflated for GPT-4o and Llama relative to the user equation. The controls for length and similarity do not address this because the confound is the category label itself. The paper should report the full confusion matrix, propagate classification uncertainty into Table 2 (e.g., by bootstrapping labels), and provide a robustness version estimated on the 3,500 hand-labeled responses. The limitation discussion in §5.4 does not currently mention this threat.","section":"§3.2, Table 2"},{"comment":"The text reports p < 0.001 for the difference between the user and GPT-4o ethical-refusal coefficients and p > 0.05 for technical refusals, but no test statistic or procedure for comparing coefficients across the separate regressions is described anywhere in §4. Separate OLS columns do not supply a cross-equation p-value; the authors should specify the test used (e.g., a Wald test on a pooled model with evaluator interactions, seemingly unrelated estimation, or a bootstrap) and report the test statistics for the ethical versus technical contrasts. Without this, the inferential claim for H1 is not auditable.","section":"§4, Table 2 and text"},{"comment":"The unit of analysis is the individual response, but responses are nested in 49,938 comparison pairs, and within a pair the win/loss outcomes are exactly complementary (and ties are shared). The reported robust standard errors are not clustered by pair or prompt, so the effective sample size for inference on the refusal coefficients is smaller than the number of response-level observations suggests. Clustering by pair (and possibly by prompt) should be reported; given the large effects this is unlikely to overturn H1, but it is needed for correct inference. The paper should also explain why the GPT-4o regression has 49,900 observations rather than 49,938.","section":"§3.1, Table 2"}],"minor_comments":[{"comment":"The phrase 'Accuracy & F1 Score = 88%' is ambiguous; please state whether this is macro-F1, weighted F1, or a single overall number, and report class-wise F1 for all five categories, especially 'Refusal Ethical', which is the category driving the main claim.","section":"§3.2"},{"comment":"The paper says the standardized prompt format from Zheng et al. (2023) is used, but does not provide the exact prompt, model temperature, max tokens, or API versions; an appendix with the precise prompt and inference settings would help reproducibility.","section":"§3.3"},{"comment":"Please report the date/version of the Chatbot Arena data used, the exact criterion for selecting 'one-turn' conversations, and how the three-way outcome (win/loss/tie) is coded in the regression (presumably two indicator outcomes with tie as the omitted reference).","section":"§3.1"},{"comment":"Since the opponent-side controls are included but not shown, report their coefficients or at least state clearly that all four opponent refusal categories are included as dummies; this would help readers assess whether the Refusal Ethical effect is net of opponent behavior.","section":"Table 2"},{"comment":"The reference list contains minor inconsistencies, such as in-text 'Saha et al., 2024' versus the reference year 2025, and 'Whittlestone et al.' appearing with both 2019 and 2021 dates; please reconcile these.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's reliance on the author's own prior classifier is a transparency and independence concern: all labels come from a single-author system with no external validation, and the prior work is referenced rather than included. The central comparison is not circular—the human-versus-judge divergence is a new empirical result—but the dependence on self-created labels makes the finding harder to verify than if an independent classifier or a hand-coded subsample were used. If the prior paper is unavailable or not reproducible, the editors should weigh whether this manuscript can stand alone as a complete contribution. The two-model replication and controlled regressions are genuine strengths, but the classifier validation gap should be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's core finding is worth knowing: both GPT-4o and Llama 3 70B evaluate ethical refusals more favorably than human users do, while the same is not true for technical refusals. That's a clean, novel result that extends the list of LLM-as-a-judge biases. The tables are readable, the regressions control for length, similarity, and opponent model, and there are two judge models from different families. I believe the result.\n\nThe soft spots are not fatal, but they are real. The biggest is the classifier. The paper takes a RoBERTa classifier from the author's prior work, reports only overall 88% accuracy, and defers the confusion matrix to that paper. The categories are imbalanced—ethical refusals are 2.6% of responses—so class-wise precision/recall matters. The stress-test note's specific claim that refusal-disclaimer confusion would \"produce exactly the observed divergence\" doesn't hold up arithmetically: mixing disclaimers into the refusal category would actually shrink the divergence, not inflate it. But the general concern stands: if label noise is differential across categories and correlated with judge preferences, the coefficient estimates could be biased. The paper needs to show the confusion matrix, report class-wise metrics, and ideally do a robustness check with a different classifier or a human re-validation subset.\n\nSecond, the abstract overstates the technical-refusal result. For Llama, technical refusals are penalized more than by users (win coefficient -0.294 vs -0.200), which is a divergence, just in the opposite direction. The paper's own Section 4 is more careful about this, but the abstract says \"a divergence not observed for technical refusals.\" That should be fixed.\n\nThird, the cross-evaluator significance tests are not described. The paper says the GPT-4o versus user difference for ethical refusals has p<0.001 and the technical difference p>0.05, but there's no method (pooled regression with interactions? bootstrap?). Without that, those claims aren't reproducible. And there's no code or data.\n\nNone of these are deal-breakers. The central argument holds up; the pattern is consistent with what alignment training would predict. This deserves a serious referee, but the referee should ask for the classifier validation and the significance-test details before acceptance.\n\nBest,\n[Your name]","headline":"A genuinely new LLM-as-a-judge bias finding that rests on a single-author classifier the paper doesn't fully validate.","tokens_in":11123,"tokens_out":4958,"would_cite":false,"duration_ms":40490,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LLM-as-a-Judge systems rate ethical refusals more favorably than human users do, a gap the paper calls moderation bias.","keywords":["LLM-as-a-Judge","content moderation","response refusals","ethical alignment","human preference","moderation bias","Chatbot Arena","evaluation bias"],"falsifier":"Re-label a random sample of the 99,876 responses by hand, re-run the user-versus-judge win/loss regressions using only the responses where the hand label matches the classifier, and check whether the ethical-refusal gap persists; if the gap shrinks to nothing on the confirmed labels, the moderation bias depends on classifier error rather than on genuine judgment differences.","tokens_in":10063,"feed_emoji":"🤖","tokens_out":8628,"duration_ms":66158,"temperature":0.7,"pith_summary":"The paper asks whether automated model judges evaluate refusal responses differently from the human users who vote on the same responses. Using 49,938 one-turn comparison pairs from a large human-preference platform, it classifies each response as an ethical refusal, technical refusal, disclaimer, or standard response, then has two language-model judges (GPT-4o and Llama 3 70B) re-evaluate the same pairs. It finds that ethical refusals—responses that decline a prompt for safety or normative reasons—win far more often under both AI judges than under human voters, while technical refusals do not show the same gap. The paper names this divergence moderation bias and treats it as evidence that alignment-trained evaluators systematically reward safety-consistent refusal behavior beyond what users prefer. If correct, the result implies that any pipeline using LLM judges for benchmarking or training will over-weight ethical refusals relative to user satisfaction.","feed_headline":"AI judges reward safety refusals four times more than users","feed_subtitle":"Model judges rate ethical refusals far more favorably than humans, tilting training rewards toward refusal.","key_machinery":"The machinery is a two-part measurement setup: a five-way classifier that sorts responses into ethical refusals, ethical disclaimers, technical refusals, technical disclaimers, and standard responses, paired with a pairwise LLM-as-a-Judge protocol in which a judge model sees the prompt and two anonymized responses and picks a winner, a loser, or a tie. The classifier is a transformer-based text classifier trained on 3,500 hand-annotated responses and applied to all 99,876 responses; the judge protocol follows a standardized prompt, with response length and prompt-response similarity included as controls. This setup lets the paper compare win/loss/tie rates for each response type across human voters and two AI judges, and the contrast between ethical and technical refusals is what isolates the bias.","core_discovery":"On the paper's own terms, the central discovery is that LLM-as-a-Judge evaluation carries a moderation bias: model judges consistently rate ethical refusals more favorably than human users, and this divergence disappears for technical refusals. In the descriptive data, ethical refusals win only 8% of human comparisons but 31% under GPT-4o and 27% under Llama 3 70B, while technical refusals move from 16% under humans to 27% and 24% respectively. The regression results show the user penalty for an ethical refusal is a 32 percentage point reduction in win rate, versus only 12 points under GPT-4o and 23 under Llama 3 70B; for technical refusals the penalties are statistically indistinguishable between users and GPT-4o. The authors interpret the pattern as alignment signaling: models trained to value safety and normative compliance reward ethical refusals as good behavior, whereas human users experience them as evasive or moralizing.","pith_inferences":["Editorial inference: The paper measures an association, not a causal effect; a direct test would compare the same base model as a judge before and after safety fine-tuning to see whether alignment training alone creates the bias.","Editorial inference: The user-vote baseline is itself a particular population, so the 'bias' partly reflects whose preferences count; a more diverse human rater pool could shrink or reshape the gap.","Editorial inference: A practical extension would be to report refusal-type-specific agreement between LLM judges and human raters on evaluation cards, making moderation bias visible in routine model audits."],"forward_implications":["If LLM judges are used as training signals, ethical refusals will be rewarded more than human users reward them, pushing deployed models toward more refusal-heavy moderation.","Benchmarks that score models with LLM judges will systematically favor safety-aligned refusal responses, so high benchmark scores can coexist with low user satisfaction.","Because the effect shows up in both a proprietary and an open judge model, it is likely a general property of alignment-trained judges rather than a quirk of one model.","Since technical refusals do not show the divergence, the bias is about ethical alignment signaling specifically, not about declining to answer in general."],"supporting_citations":[{"why":"Supplies the five-way refusal classifier and the earlier finding that users penalize ethical refusals more than technical ones.","marker":"Pasch (2025)"},{"why":"Provides the Chatbot Arena comparison-pair dataset that all user and judge evaluations are drawn from.","marker":"Chiang et al. (2024)"},{"why":"Defines the standardized pairwise LLM-as-a-Judge prompt format used for both judge models.","marker":"Zheng et al. (2023)"},{"why":"Documents known LaaJ stylistic biases, the background against which moderation bias is positioned as a new bias type.","marker":"Ye et al. (2024)"},{"why":"Establishes RLHF as the training mechanism that teaches models to value helpful and aligned outputs, the proposed cause of the bias.","marker":"Ouyang et al. (2022)"},{"why":"Supports the alignment-training account by showing how safety and harm-avoidance preferences are instilled in models.","marker":"Bai et al. (2022)"},{"why":"Justifies the choice of GPT-4o and Llama 3 70B as strong, widely used judge models.","marker":"Raju et al. (2024)"},{"why":"Provides the sentence-embedding model used to compute the prompt-response similarity control.","marker":"Reimers & Gurevych (2019)"}],"fun_headline_variants":["AI judges reward ethical refusals more than humans do","Model judges show moderation bias toward safety refusals","LLM judges rate ethical refusals higher than human users","AI evaluators favor ethical refusals, humans less so","Refusal divergence: AI judges vs human users on ethics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole user-versus-judge comparison rests on the automated classifier correctly sorting every response into ethical refusal, technical refusal, disclaimer, or standard response; if its labeling errors line up with how judges rate those responses, the moderation bias could be an artifact of misclassification rather than a real difference in judgment.","fun_headline_variants_meta":{"raw":{"variants":["AI judges reward ethical refusals more than humans do","Model judges show moderation bias toward safety refusals","LLM judges rate ethical refusals higher than human users","AI evaluators favor ethical refusals, humans less so","Refusal divergence: AI judges vs human users on ethics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2682,"prompt_tokens":1047,"completion_tokens":1635,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1553}},"tokens_in":663,"tokens_out":1635,"duration_ms":9836,"temperature":1.0,"reasoning_tokens":1553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:19:21.856947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-label a random sample of the 99,876 responses by hand, re-run the user-versus-judge win/loss regressions using only the responses where the hand label matches the classifier, and check whether the ethical-refusal gap persists; if the gap shrinks to nothing on the confirmed labels, the moderation bias depends on classifier error rather than on genuine judgment differences.","supporting_citations":[],"review_version":1}