{"id":"6939425a-7f04-4e9c-9092-abacee46ddce","arxiv_id":"2506.11047","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A pilot study suggests people can sometimes spot visible salary disparities in dot plots, but one positive cluster out of four is too little evidence to support perception-driven bias detection as reliable.","lead":"This paper proposes using crowdsourced human judgment of simple scatter plots to flag potential bias in datasets, and tests the idea in a small pilot. Only one of four demographic clusters showed both a statistically significant salary difference and high user flagging, so the headline reliability claim is not yet established.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed perception–bias correlation rests on one significant cluster out of four, with no participant counts, no multiple-comparison correction, and no user flag rates; reliability is not established.","rationale":"This paper is best read as a feasibility pilot: it builds a real web platform and collects pilot judgments, and the authors openly call the findings preliminary. The central claim, however, is stated in the abstract as a reliable correlation between non-expert visual perception and known bias cases. For that claim to hold, the pilot would need to show that user flag rates track statistical disparities across multiple cases. The evidence provided is one cluster (High_High) with p = 0.03557 and a textual assertion of high flagging, while the other three clusters' flag data is not given. Because four tests were run, the single marginal p-value is weak evidence on its own; the absence of response counts and the unspecified 'X%' threshold makes the aggregation step unfalsifiable as reported. Table 2 compounds the problem: its High_High M→M/M→F entries reproduce the salary means from Table 1, suggesting the MSE table is not a genuine cross-validation result, so the 'degradation in cross-group performance' claim is suspect. Phase 2 model outcomes are missing. I agree with the reader's weakest assumption and see no reason to change the REJECT verdict; the paper is promising but under-evidenced as submitted. My concrete test targets the correlation directly by requiring the missing data and a multiple-comparison correction.","tokens_in":8729,"tokens_out":4421,"duration_ms":37793,"concrete_test":"Request the full per-cluster user response counts (n and yes/no tallies) and the exact X% threshold from §3.3, then apply a Holm–Bonferroni correction to the four t-tests in Table 1. If the High_High p-value no longer survives the corrected threshold, or if the reported user flag rate for High_High is not higher than for the three non-significant clusters, the claimed perception–bias correlation is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim that 'perceptual signals from non-expert users reliably correlate with known bias cases,' the paper relies on the High Age–High Experience cluster: its t-test is significant (p = 0.03557, Table 1) and the text asserts high user-flag frequency. This is the only positive alignment among the four clusters in Table 1. No response counts, per-cluster flag rates, participant demographics, or the X% majority threshold from §3.3 are reported, so the reader cannot assess whether the alignment is real or coincidental. With four independent comparisons and an uncorrected threshold of p < 0.05, the chance of at least one false-positive cluster is about 18.5%, making a single marginal p-value weak evidence. The three non-significant clusters could act as negative controls, but their user flag data is absent; if users flagged them at similar rates, the claimed perception–statistics correspondence disappears. Table 2 is also internally inconsistent: for High_High, the M→M and M→F MSE entries equal the male and female salary means from Table 1 (12583.10 and 15222.83), indicating that the table likely lists means rather than genuine cross-validation errors, so the 'degradation in cross-group performance' conclusion is unsupported. Phase 2 model results promised in §3.4 are missing entirely. If the only load-bearing evidence is one uncorrected significant cluster without user-level data, the abstract's 'reliably correlate' and 'powerful, scalable proxy' are not justified by the submitted results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-phase framework for bias detection: crowdsourced binary visual judgments on deliberately minimal scatter plots of demographic clusters, followed by statistical validation (two-sample t-tests) and machine-learning cross-evaluation. A pilot deployment with university students and colleagues produced four demographic clusters; only the High Age–High Experience cluster showed a statistically significant salary difference between male and female subgroups (p = 0.03557). The authors claim that this alignment validates perception-driven bias detection and that perceptual signals from non-expert users can serve as a reliable, scalable proxy for fairness auditing.","tokens_in":9056,"tokens_out":5332,"duration_ms":47357,"significance":"The core idea is timely and potentially useful: if non-expert visual perception could flag biased data segments without sensitive labels, it would offer a low-cost pre-screening tool for fairness auditing. The paper also describes a concrete web platform and a falsifiable empirical protocol. However, the reported evidence is far too weak to support the abstract's claims of 'reliably correlate' and 'powerful, scalable proxy.' The pilot rests on one significant cluster out of four; no participant counts, per-cluster flag rates, inter-rater consistency statistics, or Phase 2 model results are reported; and Table 2 appears to contain a reporting error that undermines the machine-learning validation. The idea is worth pursuing, but the current manuscript does not establish the central claim.","major_comments":[{"comment":"The central claim rests on a single statistically significant cluster (High_High, p = 0.03557) out of four comparisons. With four uncorrected tests, the probability of at least one p < 0.05 by chance is about 18.5%, so this one marginal p-value provides weak evidence. The paper never reports the number of responses per cluster, the flag rates, the threshold X from Section 3.3, or any inter-rater consistency statistic, despite Section 4.3 claiming that responses were 'highly consistent.' Without these numbers, the statement that the significant cluster 'aligned with the high user-flag frequency' cannot be verified. Please report the full contingency data (users flagging vs. not flagging per cluster), the participant count, and a formal association test between perception and statistical disparity across all clusters, with appropriate multiple-comparison awareness.","section":"Section 4.4, Table 1 and Section 4.3"},{"comment":"The validation procedure is partly circular. Section 3.3 defines a 'calibrated' case as one in which a majority of users perceive a disparity AND the t-test returns p < 0.05. Section 4.4 then cites the High_High cluster, which satisfies both criteria, as evidence that perception correlates with known bias. This agreement is partly built into the case-selection criterion rather than being independently demonstrated. To support the perception–bias correlation claim, the authors need an external benchmark (e.g., synthetic data with known effect sizes, or a pre-registered set of independent bias labels) or at least a correlation computed over all cases without using the dual filter as the definition of ground truth.","section":"Section 3.3 and Section 4.4"},{"comment":"The MSE values in Table 2 are internally inconsistent: for each cluster, the M→M and M→F entries are identical to the male and female salary means reported in Table 1 (e.g., 12583.10 and 15222.83 for High_High). MSE is an average of squared errors and cannot equal a group mean. This strongly suggests the table lists means rather than cross-validation errors, so the conclusion of 'degradation in cross-group performance' in Section 4.5 is unsupported. Please provide the actual experimental MSE values with standard deviations, or remove the claim.","section":"Table 2 and Section 4.5"},{"comment":"Phase 2 of the proposed framework—training bias-predicting models—is described in Section 3.4, but no results are reported. No classifiers, features, accuracies, or predicted labels appear anywhere in Section 4. The abstract states that findings are 'validated through ... machine learning cross-evaluations,' but the only ML evidence is Table 2, which is erroneous as noted above. Please either include the Phase 2 results or substantially temper the abstract, conclusion, and abstract-level claims about automated screening.","section":"Section 3.4 and Section 4.5"}],"minor_comments":[{"comment":"The threshold '>X%' is never defined; specify how X was chosen and report its exact value in the pilot.","section":"Section 3.3"},{"comment":"The sentence 'Only the High Age – High Experience cluster showed a statistically significant difference ... thereby confirming the validity of perception-driven signals' overstates what a single significant result can establish; 'consistent with' would be more appropriate.","section":"Section 4.4"},{"comment":"The claim that 'responses were found to be highly consistent across repeated exposures and different users' needs quantitative support, such as Cohen's kappa, percent agreement, or per-cluster response counts.","section":"Section 4.3"},{"comment":"The underlying dataset is described only as 'real-world sources (e.g., Kaggle)'; provide the dataset name, version, and preprocessing steps so the t-tests and MSE results are reproducible.","section":"Section 5.2"},{"comment":"The notation M→M, M→F, F→F, and F→M is not defined in the caption or the text; please define these terms explicitly.","section":"Table 2"},{"comment":"Minor grammar: 'Figures 1 and 2 shows' should be 'Figures 1 and 2 show.'","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a pilot feasibility study, yet the abstract and conclusion make strong claims of reliability and scalability. The missing participant-level data, the undefined threshold, the circular validation design, and the apparent Table 2 error are all fixable in principle, but the authors must supply the actual data and reanalysis or substantially reframe the contribution. If the missing data cannot be provided or the corrected analyses do not support the perception–bias correlation, rejection would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the CAPTCHA-inspired idea is genuinely new, but the evidence in this pilot does not support the abstract's claim that perceptual signals 'reliably correlate' with bias. The load-bearing evidence is one significant cluster out of four, with no user counts, no flag rates, no multiple-comparison correction, and a Table 2 that appears to list group means rather than cross-validation MSEs.\n\nWhat's new: turning binary visual judgments into a scalable fairness pre-screening tool is a fresh application, distinct from FairVis and FairSight. The stripped visualizations and 'spot the difference' framing are clever and low-cost, and the two-phase pipeline—crowd flagging, then statistical/ML validation—is a sensible architecture. The paper also deserves credit for acknowledging in Section 5.2 that the visualizations are not controlled experiments and the participant pool is small.\n\nThe soft spots are serious and concentrated in the evidence. Section 4.4 reports a t-test for four clusters; only High_High reaches p < 0.05 (p = 0.03557). With four comparisons at an uncorrected threshold, the chance of at least one false positive is about 18.5%. That alone would be weak, but it is the only positive alignment. The three non-significant clusters could serve as negative controls, but the paper never reports user flag rates for them, so we cannot tell whether users flagged these at similar rates. Without that, the perception–bias correlation is unestablished. There are also no participant counts, no inter-rater consistency numbers, and no report of the X% majority threshold from Section 3.3.\n\nWorse, Table 2's MSE entries for M→M and M→F equal the male and female salary means from Table 1 exactly (12583.10 and 15222.83 for High_High). That looks like a reporting error—likely the table lists group means under the label MSE. The claim of 'degradation in cross-group performance' is therefore unsupported. And Phase 2 models from Section 3.4—the trained classifiers that mimic human perception—are never actually reported.\n\nWho is this for? A reader interested in human-in-the-loop fairness tools might find the framing worth discussing, but the paper does not yet deliver a working result. It is a feasibility sketch with one promising anecdote. I would not send it to peer review in this form; I would suggest the authors run a proper controlled study with synthetic data, report per-cluster flag rates and response counts, fix Table 2, and calibrate the abstract to the evidence.","headline":"CAPTCHA-style visual bias flagging is a genuinely fresh idea, but the pilot evidence—one significant cluster, missing user-level data, and a likely erroneous Table 2—does not back the abstract's reliability claim.","tokens_in":9539,"tokens_out":2557,"would_cite":false,"duration_ms":22052,"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":"This paper argues that aggregated visual judgments from non-expert users can flag bias in machine-learning data without sensitive labels, and it presents the High Age–High Experience cluster as the proof that perception, statistics, and…","keywords":["algorithmic bias","fairness in machine learning","crowdsourcing","human-in-the-loop","visual perception","bias detection","data visualization","label-efficient learning"],"falsifier":"Show a large batch of stripped-down scatter plots from clusters with known t-test results, including many null cases, and check whether user flag rates predict significance; if flag rates on non-significant clusters match those on the significant one, or if inter-rater agreement is near chance, the claimed perceptual proxy fails.","tokens_in":8539,"feed_emoji":"👁️","tokens_out":8288,"duration_ms":67772,"temperature":0.7,"pith_summary":"This paper tries to establish that ordinary people, without statistical training, can spot potential bias in machine-learning data simply by looking at stripped-down scatter plots of two groups. It proposes a CAPTCHA-style web task that collects binary similar/different judgments, aggregates them into flags, and then checks those flags against two-sample t-tests and cross-group model performance. The pilot found that the one cluster with a statistically significant salary gap between men and women ($p = 0.03557$) also drew strong visual flagging, which the authors read as evidence that visual intuition can be a scalable, label-free early-warning signal for fairness auditing. If true, data teams could screen data segments for bias before sensitive attributes or formal fairness metrics are available.","feed_headline":"Non-expert eyes can flag data bias from stripped plots","feed_subtitle":"In a pilot, visual flags aligned with the only significant salary gap, hinting at label-free fairness audits.","key_machinery":"The load-bearing object is the Fairness CAPTCHA: a lightweight web task that renders a data slice as a color-coded scatter plot with no axis labels, units, or numeric scale, and asks users a binary question such as 'Do these two groups look visually similar?' The aggregation of those binary answers into a flag, combined with a dual-filter calibration rule (a majority of users signal bias and a two-sample t-test returns $p < 0.05$), is what connects human perception to statistical ground truth.","core_discovery":"The central claim is that crowdsourced visual perception can serve as a reliable proxy for bias detection in structured data. The paper's demonstration is the High Age–High Experience salary cluster: users flagged that visualization at high frequency, a two-sample t-test showed a significant male–female mean salary difference ($p = 0.03557$), and cross-group regression models trained on one gender performed worse on the other, with the degradation most pronounced in that same cluster. The other three clusters showed no significant t-test result, so the alignment between perception and statistics rests on this single case, supported by the broader observation that cross-group model error degraded in all clusters.","pith_inferences":["The one significant cluster out of four suggests the true perception signal may be weak; a controlled study with many null clusters would be needed to rule out chance alignment.","The method likely measures visual separation of any two clusters rather than protected-class bias, so a crowd flag should be read as 'distributional separation worth checking' rather than as proof of unfairness.","A testable extension is to map flag rates against known effect sizes on synthetic plots, which would let practitioners set the majority threshold from measured human sensitivity instead of the paper's unspecified X%."],"forward_implications":["Data teams could pre-screen unlabeled dataset slices and focus expensive statistical audits on the segments that crowds flag.","The calibrated human flags can train classifiers that mimic collective visual judgment, giving automated bias screening for new data without sensitive attributes.","Fairness auditing could begin earlier in the ML lifecycle, before protected labels are collected, with human perception acting as a triage sensor.","Because question phrasing changed flag rates, any deployment of the method would need standardized wording and thresholds to make results comparable across studies."],"supporting_citations":[{"why":"Supplies the CAPTCHA-based crowd-labeling mechanic that the Fairness CAPTCHA tasks repurpose for fairness judgments.","marker":"[20]"},{"why":"Defines a formal fairness metric the approach is designed to complement when sensitive labels are unavailable.","marker":"[2]"},{"why":"Documents practitioner needs for fairness tooling, motivating the lightweight, label-efficient auditing design.","marker":"[12]"},{"why":"Shows visual analytics for discovering bias, the expert-oriented line this work extends to non-expert perception.","marker":"[18]"},{"why":"Establishes the precedent that non-expert crowd labels can power large-scale machine-learning data pipelines.","marker":"[21]"},{"why":"Provides the taxonomy of bias types used to frame why perceptual signals might catch disparities that metrics miss.","marker":"[10]"}],"fun_headline_variants":["Crowd visual cues flag salary bias without labels","Non-expert perception spots data bias in plots","Visual intuition matches statistically significant bias","Stripped plots let crowd detect hidden bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that fast binary visual judgments by a small, non-random group of students and colleagues reliably track statistical disparity in the underlying data, a link the pilot supports with only one significant cluster.","fun_headline_variants_meta":{"raw":{"variants":["Crowd visual cues flag salary bias without labels","Non-expert perception spots data bias in plots","Visual intuition matches statistically significant bias","Stripped plots let crowd detect hidden bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1420,"prompt_tokens":870,"completion_tokens":550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":493}},"tokens_in":486,"tokens_out":550,"duration_ms":4861,"temperature":1.0,"reasoning_tokens":493,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:11:22.190938+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Show a large batch of stripped-down scatter plots from clusters with known t-test results, including many null cases, and check whether user flag rates predict significance; if flag rates on non-significant clusters match those on the significant one, or if inter-rater agreement is near chance, the claimed perceptual proxy fails.","supporting_citations":[{"cited_title":"recaptcha: Human-based character recognition via web security measures","cited_arxiv_id":null,"evidence_quote":"Supplies the CAPTCHA-based crowd-labeling mechanic that the Fairness CAPTCHA tasks repurpose for fairness judgments."},{"cited_title":"Equality of opportunity in supervised learning","cited_arxiv_id":null,"evidence_quote":"Defines a formal fairness metric the approach is designed to complement when sensitive labels are unavailable."},{"cited_title":"Improving fairness in machine learning systems: What do industry practitioners need? In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems , pages 1–16","cited_arxiv_id":null,"evidence_quote":"Documents practitioner needs for fairness tooling, motivating the lightweight, label-efficient auditing design."},{"cited_title":"Fairvis: Visual analytics for discovering intersectional bias in machine learning","cited_arxiv_id":null,"evidence_quote":"Shows visual analytics for discovering bias, the expert-oriented line this work extends to non-expert perception."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"Establishes the precedent that non-expert crowd labels can power large-scale machine-learning data pipelines."},{"cited_title":"A survey on bias and fairness in machine learning","cited_arxiv_id":null,"evidence_quote":"Provides the taxonomy of bias types used to frame why perceptual signals might catch disparities that metrics miss."}],"review_version":1}