{"id":"8bb632c2-a673-4609-992e-46f2c9f9bf6d","arxiv_id":"2505.07339","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Across two vignette experiments, liberals and racial minorities rated affirmative algorithms as favorably as fair ones, while conservatives and White participants rated them as negatively as discriminatory algorithms, with beliefs about marginalization named as the source.","lead":"This paper reports two online experiments (N=1193) measuring how US laypeople judge algorithms that explicitly favor historically marginalized groups, algorithms that favor privileged groups, and algorithms that treat everyone equally in hiring and bail decisions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's causal 'source' claim is not supported by the design or analyses: marginalization beliefs were measured after algorithm ratings and never directly linked to those ratings.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: beliefs about marginalization were measured after algorithm judgments and were not manipulated, so the causal 'source' claim in the abstract is not established. My pass agrees. The paper's descriptive findings—fair algorithms preferred, discriminatory rejected, and affirmative ratings moderated by politics/race in domain-specific ways—are supported by the ANOVAs and contrasts reported in Section 4 and Appendix B. The concern is specifically about the mechanism claim, not about the existence of the effects. I do not see an internal inconsistency or a statistical error that would overturn the descriptive results. The paper already discloses the post-hoc coding and the single-coder procedure, and the reader's conditional verdict appropriately flags the overreach. Therefore the verdict should remain CONDITIONAL, with the condition being that the authors either soften the causal language or provide direct evidence linking beliefs to ratings. No new fatal concern emerged from my review.","tokens_in":35167,"tokens_out":1819,"duration_ms":22349,"concrete_test":"Re-analyze the posted data with marginalization-mention as a predictor or mediator of algorithm fairness/trust/support among participants in the affirmative condition, controlling for political leaning, racial group, gender, and study. If the association between mentioning a marginalized group and affirmative-algorithm ratings attenuates to zero after controls, the claimed 'source' explanation fails. Additionally, run a preregistered replication that measures marginalization beliefs before algorithm evaluation (or both before and after); if belief reports shift with algorithm condition, the original post-hoc measurement is contaminated and the causal direction claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, 'people have varying beliefs about who (if anyone) is marginalized, shaping their views of affirmative algorithms,' requires that marginalization beliefs precede and explain algorithm judgments. The study does not test this. Beliefs were elicited only after participants had already rated the algorithm (Section 3.2.3), so reports may reflect post-hoc rationalization or consistency-seeking rather than stable antecedent beliefs. The open-ended responses were coded by the first author alone (Section 3.3), with no reliability check. Most importantly, no analysis connects the coded beliefs to the algorithm ratings. The logistic regressions in Appendix C (Tables 22 and 23) model whether groups are mentioned as a function of political leaning, race, and gender; they do not include fairness, trust, objectivity, or support as outcomes. Thus the abstract's 'source ... shaping their views' statement is an interpretive leap beyond the reported statistics. The observed associations between identity/politics and algorithm evaluations are real, but the mechanism claim is unverified. A weaker, correlational version of the claim might survive, but the causal framing in the abstract and Section 5.3 is not earned.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports two between-participants vignette experiments (N = 1193, US Prolific sample) on laypeople's perceptions of fair, affirmative, discriminatory, and control algorithms in hiring and bail decisions. It finds that fair algorithms are rated positively and discriminatory algorithms negatively across political and identity groups. It reports moderation of affirmative-algorithm judgments by political leaning in hiring and by racial group in bail. It also reports null effects of context manipulations that reveal historical injustice, and it analyzes open-ended reports of which groups are considered marginalized. The paper claims that variation in beliefs about who is marginalized explains the observed disagreements about affirmative algorithms.","tokens_in":35324,"tokens_out":5206,"duration_ms":50808,"significance":"If the main results hold, the paper makes a useful empirical contribution to algorithmic fairness and HCI by documenting broad consensus on fair and discriminatory algorithms and by identifying political and racial cleavages in support for affirmative algorithms. Strengths include full ANOVA/ANCOVA tables with Bonferroni-corrected pairwise comparisons, a power analysis, and publicly available data and scripts. The specific 'source' claim about marginalization beliefs requires additional evidence; with that caveat, the study is informative for designers and policymakers working on affirmative algorithmic systems.","major_comments":[{"comment":"The abstract's statement that 'people have varying beliefs about who (if anyone) is marginalized, shaping their views of affirmative algorithms' and the corresponding discussion in §5.3 are not supported by the reported design and analyses. Beliefs about marginalization were elicited after participants had rated the algorithms (§3.2.3), so the data cannot distinguish antecedent beliefs from post-hoc rationalization. The open-ended responses were coded by the first author alone with no reliability check (§3.3), and the logistic regressions in Tables 22-23 model whether a group is mentioned as predicted by political leaning, race, and gender, not algorithm ratings. No analysis links the coded beliefs to fairness, trust, objectivity, or support. The authors should either add analyses that relate beliefs to algorithm judgments (e.g., include the coded groups as predictors or mediators) or reframe the claim as exploratory rather than identifying a causal 'source.'","section":"Abstract; §5.3; §3.2.3; Appendix C"},{"comment":"The claims that conservatives evaluate affirmative algorithms 'as negatively as discriminatory systems' (abstract) and that White participants 'evaluate affirmative algorithms as negatively as discriminatory algorithms' (§4.2.2) are not directly established by the reported statistics. Table 4 reports contrasts between slopes across algorithm type conditions, not simple-slope tests at the conservative end of the scale; a significant slope interaction does not imply equality of predicted values at a given political leaning. Similarly, Table 14 shows Non-White vs. White differences within the affirmative condition, but does not report a within-White contrast between affirmative and discriminatory conditions. Report simple slopes or estimated marginal means at relevant values (e.g., political leaning = -2 and the White subgroup) with Bonferroni correction, or soften the equivalence wording.","section":"§4.2.1; §4.2.2; Tables 4, 12, 14"}],"minor_comments":[{"comment":"Table 23 appears to contain a duplicated 'Observations' row, showing both 582 and 588 observations; please correct this typo.","section":"Appendix C, Table 23"},{"comment":"The figure captions state that standard errors are included but not visible due to their small values; consider plotting confidence intervals or numerical annotations so the uncertainty is interpretable.","section":"Figures 2-4"},{"comment":"The statement that 'the more liberal a participant was, the more critical they were about discriminatory algorithms concerning all measures' is not backed by a table of per-condition slopes; report these slopes and their confidence intervals.","section":"§4.2.1, bail decisions"},{"comment":"The coding of open-ended responses would benefit from an inter-coder reliability statistic on a subset of responses, especially because the categories (e.g., 'cisgender groups') are not self-evident.","section":"§3.3 and Appendix C"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the headline moderation results—liberals and racial minorities rating affirmative algorithms more favorably, fair algorithms broadly liked, discriminatory ones broadly rejected—are credible and worth taking seriously. Second, the paper's explanatory claim that beliefs about who is marginalized 'shape' algorithmic judgments is not tested. Beliefs were elicited after algorithm ratings and never statistically linked to those ratings in any analysis. That is an interpretive leap in the abstract, not a finding.\n\nWhat's new: this is the first work I know that directly measures lay attitudes toward affirmative algorithms as opposed to human affirmative action, and it does so across two domains with decent power. The null effect of the context manipulations—telling people the domain is unjust does not shift judgments—is a real contribution. The reporting is unusually transparent: full ANOVA tables, Bonferroni-corrected pairwise contrasts, and explicit disclosure that the contextualized hiring condition was added after initial analyses. Good practice.\n\nWhere it is soft: the 'source' claim is the main problem. The open-ended marginalization beliefs were collected after participants had already rated the algorithms, so they may be post-hoc rationalizations, and no mediation or even correlational analysis connects those beliefs to algorithm judgments. The logistic regressions predict who mentions a group from demographics, not algorithm ratings. The authors should either add a genuine mediation analysis or, given the measurement order, label this as a correlational hypothesis for future work. The abstract's 'shaping their views' phrase has to go.\n\nSecond, the abstract says conservatives and White participants evaluate affirmative algorithms 'as negatively as discriminatory systems.' That equivalence is not directly tested. The significant contrasts are between slopes or between racial groups. A simple-slope or simple-effect test within those subgroups would be needed; the figures suggest it, but the paper does not report the test.\n\nMinor concerns: the open-ended coding was done by the first author alone without a reliability check, and one experimental arm was added after initial analysis. Both are disclosed, and both warrant a sentence of caution.\n\nNet: this deserves a serious referee, but it needs major revision. The core empirical pattern likely holds; the causal framing is the load-bearing weakness. If the authors reframe the marginalization-beliefs claim as an association, the paper becomes a solid contribution.","headline":"The moderation results by ideology and race are credible and worth attention, but the abstract's causal 'source' claim about marginalization beliefs is not supported by the reported analyses.","tokens_in":35823,"tokens_out":3401,"would_cite":true,"duration_ms":33473,"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":"Affirmative algorithms divide laypeople along political and racial lines, and the split tracks beliefs about who is marginalized.","keywords":["affirmative algorithms","algorithmic fairness","algorithmic discrimination","systemic injustice","political ideology","racial identity","hiring decisions","bail decisions"],"falsifier":"An experiment that measures or experimentally manipulates beliefs about marginalization before participants rate an affirmative algorithm would settle the causal claim: if shifting these beliefs changes support for affirmative algorithms in the predicted direction, the paper's 'source' account is supported, whereas if pre-existing attitudes toward affirmative algorithms are unchanged by new evidence about marginalization, the source claim is undermined.","tokens_in":34972,"feed_emoji":"⚖️","tokens_out":5559,"duration_ms":48084,"temperature":0.7,"pith_summary":"This paper reports two experiments (N=1193) on how ordinary Americans judge three kinds of decision-making algorithms in hiring and bail decisions: fair ones that treat all demographic groups equally, affirmative ones that explicitly prioritize historically marginalized groups, and discriminatory ones that favor the historically privileged. The central finding is that fair algorithms are broadly approved and discriminatory ones broadly rejected by everyone, regardless of political leaning, race, or gender. Affirmative algorithms, however, split opinion: liberals and racial minorities rate them about as favorably as fair algorithms, while conservatives and the dominant racial group rate them about as negatively as discriminatory ones. The paper proposes that these divisions trace to differing beliefs about who, if anyone, is marginalized, and finds that people who name more marginalized groups are more supportive of affirmative algorithms.","feed_headline":"Affirmative AI splits opinion by politics and race","feed_subtitle":"Fair algorithms win broad approval; who you think is marginalized decides the rest.","key_machinery":"The central object is the algorithmic vignette: three between-participants algorithm types (affirmative, fair, discriminatory) defined solely by which demographic group the algorithm is said to prioritize, plus a control with no outcome information, embedded in identical hiring and bail decision scenarios. The explanatory mechanism is participants' open-ended reports of which groups they consider historically marginalized, coded into categories such as racial minorities, women/men, low-income groups, and LGBTQ+ individuals. These reports are used to show that beliefs about marginalization track the same political and identity lines as evaluations of affirmative algorithms, and the authors interpret the pattern through the moral-judgment lens that disagreement arises from differing beliefs about who is a vulnerable victim.","core_discovery":"The paper's core claim is that laypeople do not disagree about algorithmic fairness as such—they agree that equal treatment is fair and that favoring the privileged is unfair—but they disagree sharply about whether favoring the marginalized is fair. In the hiring domain, liberal participants evaluated affirmative algorithms as positively as fair algorithms, while conservative participants judged them as unfair as discriminatory algorithms; in the bail domain, non-White participants supported affirmative algorithms as much as fair ones, while White participants rejected them. The paper identifies the source of these divisions in participants' beliefs about marginalization: the more groups a person names as historically marginalized, the more they support algorithms that prioritize the marginalized, and these belief differences align with the same political and racial lines as the attitude differences. An additional finding is that explicitly telling participants that the decision-making domain is systemically unjust had no effect on their evaluations of any algorithm type.","pith_inferences":["The causal direction behind the 'source' claim is not established: beliefs about marginalization were measured only after participants had already rated the algorithms, so it is possible that the ratings shaped the reported beliefs rather than the reverse; a preregistered study that measures or primes marginalization beliefs before algorithm evaluation would settle this.","Because impersonal factual framing failed to move attitudes, the paper's own suggestion that personal experiences of discrimination might bridge the divide is the natural next test, and if personal narratives shift only some groups' views, the feasibility of 'affirmative algorithmic futures' may depend on targeted rather than universal messaging.","The authors note that hiring is competitive while bail is not; a direct test could swap the domains' salient identity axes (for instance, a gender-salient hiring context) to see whether the same belief-attitude link moves to gender."],"forward_implications":["Fair algorithms that disregard demographic groups enjoy near-universal support across political, racial, and gender lines.","Support for affirmative algorithms is not fixed: it depends on the domain (politics mattered in hiring, race in bail) and on who is doing the judging.","Providing information that a domain is systemically unjust does not move opinions about any algorithm type, including affirmative ones.","A person's list of which groups count as historically marginalized is a robust correlate of their acceptance of affirmative algorithms."],"supporting_citations":[{"why":"Defines affirmative algorithms and shows how risk thresholds could be lowered for historically overpoliced groups; supplies the core concept the studies evaluate.","marker":"[119]"},{"why":"Documents racial bias in criminal risk-prediction software; motivates the bail decision-making domain.","marker":"[4]"},{"why":"Reports an Amazon hiring algorithm biased against women; motivates the hiring domain.","marker":"[20]"},{"why":"Provides the moral-judgment theory used to interpret disagreements as differing beliefs about who is vulnerable.","marker":"[31]"},{"why":"Shows liberals and conservatives disagree on who is vulnerable to harm, the empirical basis for the marginalization-beliefs hypothesis.","marker":"[114]"},{"why":"Evidence that political leaning shapes support for affirmative action, extended here to algorithms.","marker":"[16]"},{"why":"Evidence that gender and race shape affirmative-action attitudes, extended here to algorithms.","marker":"[29]"},{"why":"Introduces the concept of 'affirmative algorithmic futures' that motivates the paper's implications.","marker":"[105]"},{"why":"Prior finding that racial groups agree on bail-algorithm features, contrasted with the race split observed here.","marker":"[34]"}],"fun_headline_variants":["Who's marginalized? That decides if affirmative AI is fair","Affirmative AI: fair if you see inequality, unfair if you don't","Fair is fair, but affirmative AI divides people","Injustice warnings don't change views on affirmative AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The study assumes that participants' open-ended statements about which groups are marginalized, collected after they had already rated the algorithms, capture the beliefs that caused those ratings rather than reasons they made up afterwards.","fun_headline_variants_meta":{"raw":{"variants":["Who's marginalized? That decides if affirmative AI is fair","Affirmative AI: fair if you see inequality, unfair if you don't","Fair is fair, but affirmative AI divides people","Injustice warnings don't change views on affirmative AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000863,"raw_usage":{"total_tokens":3720,"prompt_tokens":897,"completion_tokens":2823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2754}},"tokens_in":513,"tokens_out":2823,"duration_ms":22931,"temperature":1.0,"reasoning_tokens":2754,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:18:09.873276+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An experiment that measures or experimentally manipulates beliefs about marginalization before participants rate an affirmative algorithm would settle the causal claim: if shifting these beliefs changes support for affirmative algorithms in the predicted direction, the paper's 'source' account is supported, whereas if pre-existing attitudes toward affirmative algorithms are unchanged by new evidence about marginalization, the source claim is undermined.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines affirmative algorithms and shows how risk thresholds could be lowered for historically overpoliced groups; supplies the core concept the studies evaluate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows liberals and conservatives disagree on who is vulnerable to harm, the empirical basis for the marginalization-beliefs hypothesis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the concept of 'affirmative algorithmic futures' that motivates the paper's implications."}],"review_version":1}