{"id":"f4e4648f-0e72-42ba-8692-416541ee3d49","arxiv_id":"2411.17374","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"On one admissions dataset, ML models score higher on a consistency-based fairness metric than human decision makers, while keeping similar accuracy.","lead":"This study compares how consistently machine learning models and human evaluators treat similar university applicants, using a private dataset of 870 profiles. The authors find that ML models make more consistent decisions than humans without losing accuracy, and argue for combining both approaches.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ML models are trained on the same human decisions used to define the human consistency baseline; the reported 14–19% gap may be a smoothing artifact, not evidence that ML is fairer than humans.","rationale":"The reader's similarity-space concern is legitimate but not the decisive flaw: even if BERT embeddings perfectly capture applicant similarity, the central claim still fails if the ML advantage is an artifact of fitting smooth functions to the same human labels used as the human baseline. Section 3.1 shows the only label is Type (Offered/Not Offered); Section 3.3 uses human decision points as targets; the ML models are trained on these labels. Human consistency is computed on raw labels; ML consistency on model predictions. Fitting a regularized classifier to noisy labels typically increases Eq. (1) consistency because predictions are a smoothed function of the input features used to define neighborhoods. The paper provides no independent human decisions (multiple raters per profile) and does not state that human and ML consistency are computed on the same applicant subset. Therefore the 14.08–18.79% margin may be a numerical artifact, not a finding about human versus machine fairness. The proposed test—collecting held-out human ratings and comparing model predictions to those ratings—would settle whether the advantage survives an independent baseline. This reinforces the reader's REJECT verdict; no adjustment is needed.","tokens_in":5442,"tokens_out":12754,"duration_ms":129270,"concrete_test":"Collect independent human decisions on a held-out set of the same 870 profiles (e.g., two or more raters per profile), train XGB/Bi-LSTM only on the remaining profiles' Type labels, then compute Eq. (1) consistency on the held-out profiles for (a) each rater's decisions and (b) the model predictions, with bootstrap confidence intervals. If the model advantage shrinks to within noise, the reported gap is an artifact of smoothing the very labels used as the human baseline. If no new labels are possible, at minimum recompute human consistency on the 10% test set used for ML predictions and compare like-for-like.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec. 4.4) is that ML models achieve much higher Eq. (1) consistency than humans. But the comparison lacks an independent human-decision baseline. Section 3.1 defines the only label as Type (Offered/Not Offered); Section 3.3 treats SL/AR/OF as human 'targets'. The XGB/Bi-LSTM models are trained to predict these human-generated labels. Human C(AR)=0.5632 and C(OF)=0.6023 in Table 1 are computed directly on the raw human labels, while ML C(AR)=0.8073 and C(OF)=0.7797 are computed on predictions of a regularized function fitted to those same labels (with an 80/10/10 split, but the paper never states that human consistency is computed on the same test subset). Any smooth classifier fitted to noisy labels will tend to be more self-consistent in feature space than the raw labels, because fitting averages over label noise. Thus the 14.08–18.79% margins may quantify the gap between a smoothed reconstruction of human decisions and the unsmoothed originals, not a property of human versus machine fairness. No multiple-rater or repeated-decision data are described, so there is no evidence that 'human consistency' measures human decision-maker reliability rather than outcome noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies individual fairness in university admissions using 870 applicant profiles from an existing dataset. It introduces a consistency score (Eq. 1) based on k-nearest neighbors in BERT embedding space, applies it to human decisions at three stages (shortlisting, admission recommendation, offer) and to three ML models (KNN, XGB, Bi-LSTM), and reports that XGB and Bi-LSTM exceed human consistency by roughly 14 to 19 percentage points while maintaining comparable accuracy. The authors conclude that ML models could support a hybrid human-ML admissions pipeline by providing more consistent treatment of similar applicants.","tokens_in":5702,"tokens_out":6124,"duration_ms":61535,"significance":"If the headline result were valid, the paper would make a useful empirical contribution: it applies individual fairness to a realistic admissions task, uses actual human decision stages, and proposes a simple and transparent consistency metric in Eq. (1). The comparison of multiple model families and the use of contextual embeddings are also reasonable starting points. However, the central comparison is currently not an apples-to-apples measurement: human consistency is computed on raw human decision labels, while model consistency is computed on predictions of classifiers fitted to those same labels, so a large part of the reported gap can be explained by smoothing of label noise. The paper also lacks uncertainty quantification, omits KNN consistency scores, and does not validate the similarity space on which the fairness definition depends. These issues are load-bearing for the paper's main claim.","major_comments":[{"comment":"The central claim that ML models are more individually fair than humans is not established because the human and ML consistency scores measure different objects. Human C(AR)=0.5632 and C(OF)=0.6023 are computed directly on raw human decisions, whereas ML C(AR)=0.8073 and C(OF)=0.7797 are computed on predictions of classifiers trained to reproduce those same human labels. Any regularized classifier fitted to noisy labels smooths label noise, so its nearest-neighbor consistency in the embedding space will tend to exceed the raw labels' consistency even if the underlying decision process is identical. The paper never states that human consistency is computed on the same 10% test subset used for model evaluation, and no repeated decisions or multiple-rater data are described that could serve as an independent human consistency baseline. This is a load-bearing issue for RQ2 and for the abstract's headline margin.","section":"§3.3, §4.4, Eq. (1)"},{"comment":"The consistency score's validity depends entirely on the choice of similarity space and on k, since Eq. (1) uses k-nearest neighbors in the BERT embedding space to decide which applicants are \"similar.\" The paper does not validate this embedding-based similarity against any admissions-relevant notion of similarity, and it reports no sensitivity analysis for k or for the embedding construction. Without such validation, the reported consistency scores are not necessarily measurements of individual fairness, because a different embedding or k could mechanically change every score in Table 1.","section":"§4.3, Eq. (1)"},{"comment":"Table 1 reports no consistency scores for the KNN model, yet Section 4.4 states that \"both machine learning models demonstrate substantially higher consistency scores\" after presenting three ML models. If KNN's consistency is unavailable, all claims about ML models must be restricted to XGB and Bi-LSTM; if it is available, it should be reported, since the retrieval-based KNN classifier is directly relevant to the k-nearest-neighbor definition in Eq. (1).","section":"Table 1, §4.4"},{"comment":"All consistency and accuracy results are single point estimates from one 80/10/10 split of 870 profiles, with no error bars, bootstrap intervals, or significance tests. Because the raw classification metrics are close between humans and models (for example, Bi-LSTM accuracy 82.76% vs. human AR accuracy 80.87%), the reported 24.41 percentage-point consistency gap could be an artifact of the particular split, and the paper provides no evidence to rule this out.","section":"§4.4, §3.1"},{"comment":"The accuracy comparison in RQ3 conflates two different targets. The ML models are trained to predict the human decision labels defined in Section 3.3, so their accuracy measures agreement with human decisions, not performance on the final admission outcome Type. The human decision rows in Table 1 are also evaluated against Type, so statements such as \"Bi-LSTM ... even surpassing human decisions in accuracy\" compare model accuracy on human labels with human accuracy on final offers, which are different quantities. This undermines the claim that ML models \"outperform human in making admission decisions.\"","section":"§4.4, RQ3"}],"minor_comments":[{"comment":"The text states there is \"a gradual decrease in consistency from OF to AR stages, i.e. from 60.23% to 56.32%\" and then says biases become more pronounced in later stages; since OF is later than AR and has higher consistency, this interpretation moves in the wrong direction.","section":"§4.5"},{"comment":"The sentence \"early stages ... demonstrate strong classification performance i.e. over 8% in F1-Score and accuracy\" should say \"over 80%,\" since the reported values are 0.81 to 0.85.","section":"§4.5, Table 1"},{"comment":"The margins 14.08% to 18.79% are absolute percentage-point differences, not relative improvements; for example, 80.73% versus 56.32% is a 43.3% relative increase. The paper should phrase these as percentage points or report relative changes.","section":"Abstract, §4.5"},{"comment":"Reference 14 lists \"Unknown, A.\" as the author; this placeholder must be completed before publication.","section":"References"},{"comment":"The described vector dimension needs clarification: four features with 768-dimensional [CLS] embeddings would concatenate to 3072 dimensions, while the paper reports 3840 dimensions, presumably because the concatenated Combined document is also included as a fifth feature; this should be stated explicitly.","section":"§3.2"}],"recommendation":"reject","confidential_remarks":"The manuscript is a compact empirical study whose central claim depends on an uncontrolled comparison between raw human labels and smoothed model predictions. I would encourage the authors to resubmit after redesigning the human baseline using repeated decisions or multiple raters, reporting matching test-set consistency with uncertainty, and validating the similarity space. I also note that the dataset is cited to the authors' own prior work (ref. 8), so its availability should be verified during any future review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's headline result—that ML models are 14–19% more consistent than human evaluators—probably doesn't survive contact with its own methodology. The models are trained to predict the human decisions (Section 3.3), and their consistency scores are computed on those fitted predictions. Human consistency is computed on the raw labels. A regularized model fitted to noisy labels will always look smoother than the labels themselves, so the reported gap may just be the difference between a smoothed reconstruction and the original noise. The paper never tells us whether human consistency is computed on the same ten-percent test split used for the ML scores, which makes the comparison apples-to-oranges even in principle.\n\nThat said, the paper has real virtues. It applies a sensible, established consistency metric (Zemel et al. 2013) to a real admissions dataset, and it is clearly written. The idea of measuring individual fairness as local smoothness in a feature space is appropriate, and the accuracy numbers for the ML models are reported transparently. The empirical exercise is straightforward, and the authors are honest about the lack of code/data.\n\nThe soft spots are load-bearing, not minor. The lack of an independent human baseline is the central flaw; without repeated decisions by multiple raters, \"human consistency\" is just outcome noise. The title promises a fairness-accuracy trade-off analysis, but the paper merely lists both metrics; no trade-off is actually modeled or tested. Error bars and significance tests are missing throughout, so we have no sense of whether the 14–19% margin is even statistically distinguishable from zero. The KNN model's low accuracy (68.97%) muddies the comparison, and the embedding-based similarity space is not validated against any human notion of applicant similarity.\n\nThis is a workshop-level paper with a good question and a flawed answer. It could be useful as a cautionary example of how not to compare human and ML decision consistency, but it is not evidence that ML is fairer than humans. A serious referee would likely ask for major revisions or reject it outright, and I would not want to cite it as a factual result.\n\nRecommendation: do not accept for peer review as is; reject, but encourage the authors to redo the analysis with multiple raters and a proper held-out evaluation.","headline":"A plausibly written paper whose central ML-vs-human consistency comparison is likely a smoothing artifact, since the ML models are trained on the very human labels they are compared against.","tokens_in":6219,"tokens_out":2343,"would_cite":false,"duration_ms":24124,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Machine learning models make more consistent admissions decisions than human experts.","keywords":["individual fairness","consistency score","university admissions","BERT embeddings","machine learning fairness","human decision-making","fairness-accuracy trade-off","Bi-LSTM"],"falsifier":"If a sample of human evaluators, shown pairs of applicants selected as k-nearest neighbors by the BERT embeddings, do not rate them as similarly qualified, then the consistency score ceases to be a measure of individual fairness as humans understand it, and the ML advantage could be an artifact of the embedding space.","tokens_in":5243,"feed_emoji":"⚖️","tokens_out":6788,"duration_ms":53217,"temperature":0.7,"pith_summary":"The paper asks whether promoting fairness in machine-learning admission models must come at the cost of accuracy, and answers no: it measures individual fairness with a consistency score and finds that the XGB and Bi-LSTM models separate similar applicants more consistently than human evaluators do at three decision points. On a real dataset of 870 university applicant profiles, Bi-LSTM reaches 80.73% consistency on admission recommendations and 77.97% on offers, against human scores of 56.32% and 60.23%. The paper reports that ML consistency exceeds human consistency by margins of 14.08% to 18.79%, and argues that ML models can serve as decision-support tools to flag inconsistent human judgments while keeping accuracy at human-comparable levels.","feed_headline":"ML models beat human experts on admissions consistency","feed_subtitle":"On 870 real applicant profiles, Bi-LSTM hits 80.7% consistency versus 56.3% for human reviewers, at comparable accuracy.","key_machinery":"The consistency score, defined as $C = 1 - \\frac{1}{N}\\sum_{i=1}^N \\left| \\hat{y}_i - \\frac{1}{k} \\sum_{j \\in \\text{knn}(i)} \\hat{y}_j \\right|$, measures how similarly a classifier treats each applicant and its $k$ nearest neighbors, with $k=5$. The nearest neighbors are computed in a BERT embedding space: each applicant's four textual features are encoded with bert-base-uncased, the [CLS] token gives a 768-dimensional vector per feature, and concatenation yields a 3840-dimensional profile vector from which a feature-reranked similarity matrix is built. This similarity matrix is the machinery that makes the fairness comparison between humans and models possible, because it fixes a common notion of 'similar applicant' for both.","core_discovery":"The central claim is that individual fairness, measured as decision consistency among similar applicants, is substantially higher for machine learning models than for human admission reviewers, and this fairness gain does not hurt accuracy. On the 870-profile admissions dataset, Bi-LSTM attains 80.73% consistency for admission recommendations and 77.97% for offers, compared with human scores of 56.32% and 60.23%, while its accuracy (82.76%) actually exceeds human accuracy at those stages (80.87% at AR). The paper interprets the human results as evidence of cognitive bias that grows in later decision stages, and proposes a hybrid pipeline where ML consistency checks prompt re-review of divergent cases.","pith_inferences":["The general claim likely depends on the specific dataset and embedding; we would not extrapolate to other domains such as hiring or lending without similar experiments, since the BERT similarity space may not match each domain's fairness-relevant features.","The consistency metric as defined treats predictions as binary and averages over nearest neighbors; a finer-grained ordinal decision scale or a fairness definition based on ground-truth outcomes rather than predicted labels would likely change the absolute scores, though the relative human-model gap might persist.","A practical testable extension is to run the same consistency measurement on a dataset several times larger to see whether the ML advantage shrinks as human reviewers gain more experience with the applicant pool."],"forward_implications":["If ML models are more consistent than humans at similar profiles, using them as an initial screening stage in admissions could reduce unwanted variability without sacrificing accuracy.","Disagreements between a model's decision and a human reviewer's decision can be flagged for additional review, targeting exactly the cases where human bias is most likely.","The consistency score gives admissions offices a quantifiable fairness metric beyond accuracy, letting them audit decision processes over time.","The finding that human consistency drops from shortlisting to later stages suggests biases enter when decision criteria become more complex, so interventions should focus there."],"supporting_citations":[{"why":"Supplies the real-world admissions dataset of 870 applicant profiles with human decision labels used for all model and human comparisons.","marker":"[8]"},{"why":"Defines the consistency score formulation that quantifies individual fairness as agreement among k-nearest neighbors.","marker":"[17]"},{"why":"Provides the individual fairness principle that similar individuals should receive similar treatment, which the consistency score operationalizes.","marker":"[6]"}],"fun_headline_variants":["AI admissions fairer than human reviewers, accuracy intact","Fairness gain without accuracy loss in machine admissions","ML models outscore human experts on fair admissions","Consistent AI beats human bias in college admission calls"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The consistency score assumes that similarity among applicants is accurately captured by distances in a BERT embedding space, so that the k-nearest neighbors in that space really are the 'similar individuals' whose treatment should match.","fun_headline_variants_meta":{"raw":{"variants":["AI admissions fairer than human reviewers, accuracy intact","Fairness gain without accuracy loss in machine admissions","ML models outscore human experts on fair admissions","Consistent AI beats human bias in college admission calls"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000375,"raw_usage":{"total_tokens":1947,"prompt_tokens":837,"completion_tokens":1110,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":1049}},"tokens_in":453,"tokens_out":1110,"duration_ms":8539,"temperature":1.0,"reasoning_tokens":1049,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:11:35.661533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a sample of human evaluators, shown pairs of applicants selected as k-nearest neighbors by the BERT embeddings, do not rate them as similarly qualified, then the consistency score ceases to be a measure of individual fairness as humans understand it, and the ML advantage could be an artifact of the embedding space.","supporting_citations":[{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world admissions dataset of 870 applicant profiles with human decision labels used for all model and human comparisons."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Defines the consistency score formulation that quantifies individual fairness as agreement among k-nearest neighbors."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Provides the individual fairness principle that similar individuals should receive similar treatment, which the consistency score operationalizes."}],"review_version":1}