{"id":"17ade127-e51d-48b8-973b-eae0e2810c09","arxiv_id":"2507.21081","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"People choose explanations partly to spare listeners' regret, and a regret-aware computational model captures this behavior.","lead":"Researchers modeled how doctors choose what to tell patients, adding a term for the patient's potential regret. The model matched human choices better than versions that ignored emotion, suggesting some people deliberately soften explanations to protect listeners.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The regret term's success is not specifically validated: c_social may just track a simpler norm of not blaming insecure patients, so the emotion-specific conclusion is not yet established.","rationale":"The reader's weakest assumption identifies the same load-bearing point: c_social is asserted to track regret but is not independently validated. My read agrees and adds that the model's capacity to fit temperament-dependent avoidance of mentioning drinking could mimic regret without any counterfactual emotion computation. The paper's own admission that the causal-structure manipulation was collapsed and that the model uses a simplified probability-of-disease rule further weakens confidence in the exact counterfactual computation, but that is secondary. The evidence is still consistent with the conclusion under a broader reading, and the milk replication partially supports the general framework, so I would not reject the paper. The conditional verdict already captures the missing validation; my concern strengthens the condition but does not move to a different verdict.","tokens_in":10414,"tokens_out":5949,"duration_ms":65966,"concrete_test":"Fit the same tactful-participant data with a nested model in which c_social(u) is replaced by a binary blame-avoidance cost equal to 1 if the doctor mentions the drinking factor to an insecure patient and 0 otherwise, with α_social per temperament as in the full model. Compare models by leave-one-scenario-out predictive log-likelihood (or BIC) under matched parameter counts. If the blame-avoidance model matches or beats the regret model, the claim that people specifically compute felt regret is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central inference is that the full model's improvement over the no-regret ablation (r²=0.869 vs 0.58 for tactful participants) shows that people compute the patient's counterfactual regret. This requires that c_social(u)=Pr(S=1|do(D=0),u) actually tracks the felt regret driving the behavior. The paper provides no validity check for this mapping. The term is highly collinear with a much simpler social consideration: whether the explanation reveals to an insecure patient that their own past drinking was a cause of their fatal disease. A doctor who wants to avoid blaming the patient, preserve the patient's self-image, or maintain a comfortable clinical relationship would produce the same utterance choices without representing regret as a counterfactual utility. Moreover, since α_social is fit separately for confident and insecure patients, the model can absorb any temperament-dependent reluctance to mention drinking; the specific counterfactual form of c_social is not tested against that simpler alternative. The milk replication weakens the point further: for tactful participants, α_social for R=1 is no longer significantly different from zero, so the only direct evidence for the regret mechanism comes from a single alcohol study.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether explainers take listeners' emotions into account when crafting explanations. It extends Chandra et al.'s rational-communication framework by adding a social-cost term that penalizes the listener's counterfactual regret, defined as c_social(u) = Pr(S=1 | do(D=0), u). In a Prolific experiment, participants playing a doctor chose whether to mention a virus and a drinking-threshold factor when explaining a fatal disease to confident versus insecure patients. Participants were split by self-report into 'tactful' and 'candid' groups. The full model with separate regret weights for confident and insecure patients fits tactful participants well (R²=0.869) and significantly better than an ablation without the regret term (R²=0.58), while for candid participants the regret weights are near zero. A replication replacing alcohol with milk partially reproduces the findings. The authors conclude that people do indeed reason about emotion when giving explanations. However, all model parameters are fit in-sample, the regret term is not validated as tracking felt regret rather than a simpler blame-avoidance norm, and the causal-structure manipulation was collapsed after 20% of participants failed the comprehension quiz.","tokens_in":10717,"tokens_out":5680,"duration_ms":58521,"significance":"If the central mechanism were established, this would be a useful contribution to computational accounts of explanation: it extends rational-communication models to affective listener states, connects to appraisal-based emotion models, and includes a replication in a second domain. The paper's strengths are its clean experimental design, direct model comparisons via likelihood-ratio tests and bootstrap confidence intervals, and its honest acknowledgment that the milk replication shows a weaker regret effect. The main limitation is that the key social-cost term is not construct-validated, and the reported R² and LRT values are in-sample fits rather than predictive tests. As it stands, the paper demonstrates that a model with a regret-like cost term fits better than an ablation; it does not yet establish that people compute the specific counterfactual regret quantity. With additional validation and out-of-sample checks, the contribution would be solid and well within the scope of the journal.","major_comments":[{"comment":"The paper repeatedly says the model 'predicts' human intuitions, but all six free parameters (plus the disease-probability weights) are fit to the same data on which R² and LRT statistics are computed. The likelihood-ratio tests and bootstrap confidence intervals therefore measure in-sample fit, not predictive accuracy. The claim that the full model wins because it includes regret is supported only as a nested-model comparison on the training data. To support the word 'prediction,' please add k-fold or leave-scenario-out cross-validation, or reframe the claims as model fitting and comparison. Also, the Abstract reports R²=0.87/0.97 and Figure 3 reports 0.869/0.969, while Table 3 reports 0.84/0.96 for the same full model; the discrepancy between full-data R² and bootstrap-subsample R² should be explained in the text.","section":"Abstract; §Model fit, Eq. (3), Tables 1-3"},{"comment":"The paper equates the patient's regret with c_social(u) = Pr(S=1 | do(D=0), u), following Houlihan et al. It provides no validation that this counterfactual probability tracks felt regret. A doctor who simply wants to avoid blaming an insecure patient, preserve the patient's self-image, or maintain a comfortable clinical relationship would make the same utterance choices without computing a counterfactual utility. Because α_social is fit separately for confident (R=0) and insecure (R=1) patients, the model can absorb any temperament-dependent reluctance to mention drinking; the specific counterfactual form of c_social is not tested against that simpler alternative. Please add a control model with a non-counterfactual blame-avoidance cost (e.g., a term that penalizes mentioning the drinking threshold when R=1) and compare fits, or provide independent evidence, such as emotion ratings, that c_social tracks felt regret.","section":"§Computational model, Eq. (3), c_social definition"},{"comment":"The causal-structure manipulation (conjunctive vs. disjunctive) was collapsed after 20% of participants failed the quiz, and in the model the disease probability is set to epsilon/0.25/0.5 based only on the number of present factors, independent of whether the structure is conjunctive or disjunctive. As a result, the 12 reported scenarios reduce to 6 effective cells (temperament × test result), and the paper does not actually test any claim about conjunctive versus disjunctive causal reasoning. This does not invalidate the main model comparison, but the discussion and abstract should not imply that the causal-structure factor was successfully manipulated, and the modeling assumption about disease probabilities should be flagged as a post hoc choice rather than a tested hypothesis.","section":"§Experiment; §Results"},{"comment":"In the milk replication, the LRT comparing the full model to the no-regret ablation is significant for tactful participants (Table 4, 22.84), but the fitted α_R=1_social has a 95% bootstrap CI (−0.02, 13.78) that includes zero (Table 6). Thus the only clean parameter-level evidence that the regret weight is positive for insecure patients comes from the alcohol study. The paper acknowledges this in the text, but the Abstract and Discussion still conclude broadly that 'people do indeed reason about emotion when giving explanations.' Please temper the general conclusion, or report a combined analysis across both studies that formally tests whether the regret weight is positive while accounting for domain differences.","section":"§Replication with milk, Tables 4-6"}],"minor_comments":[{"comment":"The labels 'T actful' and 'T actful participants' contain a stray space; they should read 'Tactful' and 'Tactful participants.'","section":"Figures 2 and 3"},{"comment":"The disease-probability parameter epsilon is never given a numerical value. Please state the value used in the fits.","section":"§Computational model; §Results"},{"comment":"The priors are reported as log Pr(V) and log Pr(T). Please state whether these are natural logs and how the reported values map to probabilities, so readers can reproduce the parameter values.","section":"Table 1 and Table 6"},{"comment":"The text says 125 participants were recruited and 25 failed the quiz, leaving 100; please state the final N explicitly after exclusions and clarify whether the 49%/51% split is computed on the remaining 100 participants.","section":"§Experiment, Procedure"},{"comment":"The manual reclassification of three participants based on their textual justifications could introduce experimenter bias; a blinded or pre-specified classification rule would be more robust. Please describe the procedure in more detail.","section":"§Experiment, Procedure"},{"comment":"The manuscript does not include a data or code availability statement. Given that the model is implemented in the memo PPL and the comparisons depend on fitting details, releasing code and de-identified data would aid reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision. The paper is competently executed and the model comparison is suggestive, but the central construct-validity issue for c_social and the in-sample nature of the 'predictions' need to be addressed before the claim about emotion-specific reasoning is fully supported. I would also encourage the editor to request an explicit data/code availability statement, since the modeling pipeline is non-trivial to reproduce without the memo code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core finding is solid: people who self-identify as tactful do adjust their explanations to spare an insecure patient's feelings, and the model needs a regret term to capture that. The behavioral split between tactful and candid participants is real, and the model fits the tactful group far better with the regret term than without (r² = 0.87 vs 0.58). That is a genuine new empirical result, and the milk replication is a reasonable generality check, even if it comes out weaker.\n\nWhat's actually new here is the application of the Chandra et al. rational-communication framework to an emotionally charged doctor-patient setting, with the social term reinterpreted as a listener-regret cost. The paper is honest about the causal-structure manipulation failing—20% of participants flunked the quiz and the data were collapsed—and it flags the replication's null alpha for insecure patients. That transparency earns credit.\n\nThe soft spots, in order. First, all model fits are in-sample. The six free parameters are fit to the same data used for r² and LRT, and the milk study is a second fit, not a prediction. The headline numbers therefore overstate predictive power. Second, the regret term c_social = Pr(S=1|do(D=0), u) is not validated. It is highly collinear with a simpler social norm: don't blame an insecure patient for their past drinking. Because alpha_social is fit separately for confident and insecure patients, the model can absorb exactly that temperament-dependent reluctance. The stress-test note makes this point well, and I think it holds. Third, no data or code are provided, which makes the model comparison hard to check independently.\n\nNone of this sinks the paper. The central finding—that some people avoid causing regret—is robust in the alcohol study and plausible in the milk study. What is missing is a direct test against a non-regret social-utility alternative, plus an out-of-sample or preregistered replication.\n\nFor anyone working on computational explanation, social pragmatics, or patient communication, this is worth a careful read. It deserves peer review; a serious reviewer will push for the alternative-model comparison and for artifacts. I would send it out, not desk-reject.","headline":"A clean behavioral demonstration that explanation choices are sensitive to listener regret, though the specific counterfactual-regret mechanism is not distinguished from simpler blame-avoidance.","tokens_in":11181,"tokens_out":2395,"would_cite":true,"duration_ms":26544,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that people craft explanations to manage the listener's emotion, specifically regret, and supports it with a fitted computational model.","keywords":["explanation","regret","emotion","rational communication","doctor-patient communication","counterfactual appraisal","social cognition","empathy"],"falsifier":"A cleaner version of the experiment's causal-structure contrast would settle it: in a disjunctive disease the virus alone can cause illness, so the model predicts tactful explainers should feel little need to hide the drinking cause (regret would be low even if it were mentioned). If tactful participants suppress the drinking cause just as strongly there as in conjunctive diseases, the regret mechanism is not the active ingredient.","tokens_in":10235,"feed_emoji":"💬","tokens_out":8770,"duration_ms":82347,"temperature":0.7,"pith_summary":"This paper tries to show that people craft explanations not only to transfer information but to manage the listener's emotions, specifically regret. It extends a cooperative-communication model of explanation with a social cost equal to the counterfactual probability that a patient would have avoided disease had they not drunk, then tests the model on how a doctor should explain a fatal illness. The model fits human judgments well: for tactful participants, the full model reaches $r^2 = 0.87$ while the same model without the regret term drops to $0.58$. The paper concludes that some people genuinely reason about a listener's emotional state when choosing what to explain, a finding with stakes for medical communication and AI explanation systems.","feed_headline":"Tactful explainers avoid blame to soften the listener's regret","feed_subtitle":"Model adds the patient's counterfactual regret to an explainer's goals and predicts tactful choices far better than emotion-free versions.","key_machinery":"The central object is the explainer's utility function over possible utterances, $V(u) = \\alpha_{\\mathrm{explanandum}} V_{\\mathrm{explanandum}}(u) + \\alpha_{\\mathrm{latents}} V_{\\mathrm{latents}}(u) - \\alpha_{\\mathrm{social}} c_{\\mathrm{social}}(u)$, with choices over the four possible explanations made by a softmax. Here $V_{\\mathrm{explanandum}}$ measures how well the explanation resolves the patient's expectation violation, $V_{\\mathrm{latents}}$ penalizes false inferences about unmentioned causes ('lying by omission'), and $c_{\\mathrm{social}}(u) = \\Pr(S=1 \\mid \\mathrm{do}(D=0), u)$ is the counterfactual probability of the disease had the patient abstained from alcohol — the term that operationalises regret. The paper fits this machinery to data separately for tactful and candid participants, and the fitted regret weight is what carries the empathic behaviour: large for insecure patients under tactful participants, indistinguishable from zero otherwise.","core_discovery":"People do indeed reason about emotion when giving explanations. In a doctor–patient vignette, participants who described themselves as tactful suppressed a true but blameworthy cause (excessive drinking) when the patient was insecure about past choices, while candid participants reported all causes regardless of temperament. A computational model in which the explainer maximises a weighted utility — understanding, correct inference about unmentioned causes, and a penalty proportional to the patient's counterfactual regret $\\Pr(S=1 \\mid \\mathrm{do}(D=0), u)$ — reproduces these choices: for tactful participants the fitted regret weight is large for insecure patients and near zero for confident ones, and the full model beats every ablation, most tellingly the no-regret version ($r^2 = 0.87$ vs $0.58$). The paper treats this as evidence that explanation is a cooperative social act sensitive to the listener's affective state, and replicates the pattern with milk instead of alcohol.","pith_inferences":["A behavioural hidden-variable model could infer the tactful/candid split from choice data alone, removing the reliance on self-reports and testing whether the two groups are genuinely distinct or ends of a continuum.","The regret term conflates self-focused regret with other-focused blame: mentioning college drinking might trigger shame or perceived judgment, not only counterfactual regret. A variant that separates these would clarify which emotional cost actually shapes explanations.","The same machinery predicts the reverse phenomenon: an explainer may volunteer an external cause even when the true cause is internal, if doing so spares the listener regret — a form of face-saving that goes beyond the doctor vignette.","The weaker milk replication suggests the social cost should be scaled by the moral valence or stability of the act (alcohol is more blameworthy than milk); a parametric version of $c_{\\mathrm{social}}$ that factors in blameworthiness would make the model more portable."],"forward_implications":["Medical-communication guidance should treat cause disclosure as an emotionally consequential choice, not merely an information-transfer problem.","Explainable-AI systems that ignore the user's emotional appraisal of an explanation will systematically fail to match what human explainers do.","The tactful/candid split implies that no single explanation policy fits all explainers; systems built to explain should infer or adapt to the explainer's own social-utility weights.","Swapping the regret appraisal variable for another emotion's appraisal (guilt, shame, pride) yields a concrete, testable family of models for when explainers will suppress or volunteer causes.","Because the model omits the listener's higher-order reasoning about why a cause was not mentioned, its predictions may break when the patient can infer omission; adding a level-2 patient is the paper's own stated next step."],"supporting_citations":[{"why":"The cooperative-explanation-as-rational-communication framework that this paper generalises with an emotion term.","marker":"Chandra, Chen, et al. (2024)"},{"why":"Defines the counterfactual appraisal quantity used as the regret cost $c_{\\mathrm{social}}(u)$.","marker":"Houlihan et al. (2023)"},{"why":"Shows how emotional outcomes can be planned over a theory of mind, grounding the idea of an emotional intervention term.","marker":"Chen et al. (2024)"},{"why":"Supplies the softmax pragmatic-choice model that maps utility to utterance probabilities.","marker":"Frank & Goodman (2012)"},{"why":"Gives the conjunctive/disjunctive causal-structure and normality contrasts used to build the experimental diseases.","marker":"Icard et al. (2017)"},{"why":"Establishes the listener-goal sensitivity of explanations that motivates extending social utility to emotion.","marker":"Kirfel et al. (2024)"}],"fun_headline_variants":["Explanations tailor blame to soften regret","Tactful doctors hide cause to spare regret","Empathy shapes what we explain to others","Modeling regret-aware explanation choices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on identifying the patient's felt regret with the counterfactual probability the disease would not have occurred had they not drunk; if that quantity does not track the emotion, the fitted regret term could instead be measuring blame avoidance, self-image protection, or some other motive.","fun_headline_variants_meta":{"raw":{"variants":["Explanations tailor blame to soften regret","Tactful doctors hide cause to spare regret","Empathy shapes what we explain to others","Modeling regret-aware explanation choices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1251,"prompt_tokens":853,"completion_tokens":398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":344}},"tokens_in":469,"tokens_out":398,"duration_ms":4628,"temperature":1.0,"reasoning_tokens":344,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:55:04.034146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A cleaner version of the experiment's causal-structure contrast would settle it: in a disjunctive disease the virus alone can cause illness, so the model predicts tactful explainers should feel little need to hide the drinking cause (regret would be low even if it were mentioned). If tactful participants suppress the drinking cause just as strongly there as in conjunctive diseases, the regret mechanism is not the active ingredient.","supporting_citations":[{"cited_title":", Kleiman-Weiner, M","cited_arxiv_id":null,"evidence_quote":"Defines the counterfactual appraisal quantity used as the regret cost $c_{\\mathrm{social}}(u)$."}],"review_version":1}