{"id":"78a85be1-1c31-4542-847a-dca913624573","arxiv_id":"2501.09014","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Stable Diffusion 2, XL, and 3 all produce starkly male-dominated and ethnically skewed images of software engineers, with SD3 shifting from White to Asian dominance while still suppressing Black and Arab representation.","lead":"Three versions of Stable Diffusion generate images of software engineers that are overwhelmingly male and heavily skewed toward White or Asian figures, with Black and Arab figures severely underrepresented. This case study quantifies the bias embedded in a widely used open-source image model family, giving practitioners and fairness researchers concrete evidence for mitigation work.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ethnicity underrepresentation claims rest on BLIP labels whose per-class recall for Black/Arab is never reported; aggregate accuracy and weighted F1 can mask systematic errors on rare classes.","rationale":"The reader's weakest assumption was that BLIP's automated gender and ethnicity labels are accurate enough to support the aggregate bias measurements. I agree with that identification but sharpen it: the real vulnerability is not gender (where even a 10 percentage-point systematic error cannot overturn the 79.5–99.8% male findings) nor the overall ethnicity accuracy, but per-class recall for rare ethnicity categories. A labeler can have high Accuracy and Weighted F1 while having near-zero recall for Arab or Black because those classes are a small fraction of the data. This is directly relevant to the strongest claims about underrepresentation, and it is not addressed by the paper's reported validation metrics. The gender bias finding and the SE-versus-General increase are robust to plausible labeling noise and are supported by external literature, so I would not reject or downgrade the paper. The appropriate outcome remains CONDITIONAL: the authors should supply per-class confusion matrices or an independently re-labeled sample to confirm that the Black/Arab underrepresentation is in the generated images, not in the labeler. Since this is exactly the kind of additional evidence a conditional verdict should request, the reader's verdict stands without change.","tokens_in":14480,"tokens_out":6962,"duration_ms":80632,"concrete_test":"Using the published replication package, recompute the ethnicity analysis as follows: (1) from the 534 manually labeled validation images, construct the full 5x5 confusion matrix of BLIP labels versus manual labels; (2) report per-class precision and recall for Arab and Black, with exact binomial confidence intervals, not just Weighted F1; (3) if per-class recall for either class is below roughly 0.85, apply a matrix-based correction (invert the confusion matrix) to the 6,720-image BLIP labels and re-plot Figure 2; alternatively, hand-label a new sample stratified to contain at least 50 Arab and 50 Black images and compare the resulting proportions. If the corrected or re-labeled Black/Arab proportions remain far below the 20% statistical-parity baseline with confidence intervals excluding parity, the underrepresentation claim stands; otherwise it should be materially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central ethnicity claim — 'all models significantly under-represent Black and Arab figures' — depends entirely on BLIP VQA labels aggregated in Figure 2, where the reported Black/Arab proportions are very small (e.g., SDXL SE: Black 0.91%, Arab 0.00%; SD3 General: Black 2.04%, Arab 0.89%). The validation in Section III-B and Table II reports only overall Accuracy and Weighted F1 for each SD-version/prompt-style cell. These aggregate metrics are dominated by the majority classes (White/Asian) and cannot detect systematic labeler errors on rare categories: a classifier that never outputs 'Arab' can still achieve 0.94 accuracy if Arabs are only a few percent of the sample. Because BLIP is itself a web-trained model with known demographic biases, systematic confusion (e.g., Arab misread as White, Black misread as Asian) is a real risk. The manual validation used only 89 images per cell, yielding very few Black/Arab instances and hence extremely wide per-class confidence intervals. Without a confusion matrix or per-class precision/recall for the five ethnicity categories, the measured underrepresentation of Black and Arab figures could be an artifact of the labeler rather than of Stable Diffusion. This concern does not threaten the gender finding or the broad White/Asian pattern, but it is load-bearing for the 'all models under-represent Black and Arab' part of the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports an empirical audit of gender and ethnicity bias in three Stable Diffusion versions (SD 2, SD XL, SD 3). The authors generate 6,720 images from 56 software-engineering tasks using two prompt styles, one including the keyword \"Software Engineer\" and one without role specification. They label the images automatically with BLIP VQA, validate the labels on a 534-image subsample, and compute statistical-parity style bias scores. The paper reports that all three models are strongly male-biased when depicting software engineers, that including the \"Software Engineer\" keyword increases gender bias in all versions, that SD 2 and SD XL favor White figures while SD 3 tends to favor Asian figures, and that Black and Arab figures are under-represented across all models and prompt styles. Task-level bias results and practitioner recommendations are also provided.","tokens_in":14737,"tokens_out":4850,"duration_ms":51824,"significance":"The study addresses a timely and socially important question, and its basic empirical design is a strength: 6,720 generated images, 20 repetitions per prompt, three model families, and a public replication package are concrete assets. The gender-bias finding is likely robust given the near-perfect gender-label agreement reported in Table I. If the ethnicity underrepresentation claim were fully supported, the paper would be a valuable contribution to fairness research in software engineering. However, the Black/Arab part of the headline is currently under-evidenced because the labeler validation does not report per-class accuracy on these rare categories, and the \"significance\" language is not backed by inferential statistics. These issues are fixable within the manuscript's scope.","major_comments":[{"comment":"The validation of BLIP's ethnicity labels reports only overall Accuracy and Weighted F1 per model/prompt-style cell. The central claim that \"all models significantly under-represent Black and Arab figures\" depends on very small proportions (e.g., SD XL SE: Black 0.91%, Arab 0.00%; SD 3 General: Black 2.04%, Arab 0.89%). A labeler that never outputs Black or Arab can still achieve the reported aggregate scores because these classes are only a few percent of the sample. Without per-class precision and recall, a confusion matrix, or manual verification of all images labeled Black/Arab plus a sample of false negatives, the measured underrepresentation may be an artifact of BLIP rather than of Stable Diffusion. Please report per-class validation metrics and, if possible, manually inspect all rare-class images.","section":"Section III-B, Table II, Figure 2"},{"comment":"The abstract and the RQ1/RQ2 answers use the word \"significantly\" (e.g., \"significantly increases the bias towards images representing Male figures\"), but no statistical test or confidence interval is reported for the bias estimates or for the prompt-style variations (+69%, +108%, +35%). These are point estimates from 1,120 images per cell. With these sample sizes, binomial confidence intervals are straightforward to compute. Add confidence intervals for the proportions and for the differences between General and SE conditions, or a formal test such as a chi-square or bootstrap test, and reserve \"significant\" for results supported by such evidence.","section":"Section IV-A, Tables III and IV"},{"comment":"The thresholds used to classify task-level bias as \"significantly high\" (≥0.8) and \"fair\" (≤0.2) are introduced without justification or sensitivity analysis, and they directly determine the RQ3 answer that no task has a fair ethnicity distribution. The paper should justify these thresholds, provide a sensitivity analysis, or report the continuous bias values so that readers can judge the task-level claims independently of arbitrary cutoffs.","section":"Section III-C, Table V"}],"minor_comments":[{"comment":"In the description of Cochran's formula, p is described as \"the confidence level (95% in our case)\", but in the standard formula p denotes the expected proportion. Please correct this notational confusion.","section":"Section III-B, Eq. (1)"},{"comment":"The \"±0.1\" entries appear to be the pre-specified 10% margin of error rather than estimated confidence intervals for accuracy; they produce intervals outside [0,1] for values like 0.98 and 1.00. Clarify what these intervals represent or replace them with actual bootstrap or exact confidence intervals.","section":"Tables I and II"},{"comment":"The phrase \"To account for the non-stochastic behavior of image-generation models\" should read \"stochastic behavior\"; the models are stochastic, which is why 20 repetitions per prompt are generated.","section":"Section III-A"},{"comment":"The text contains a typo, \"Perfoms support tasks\", while Table V lists \"Performs support tasks\"; please unify the wording.","section":"Section IV-C"},{"comment":"The statement that no image was mapped to the \"Other\" category would be more convincing if the paper explained how ambiguous BLIP labels were mapped to the five census-based categories and whether the two manual annotators agreed on this mapping.","section":"Section III-B, ethnicity mapping"},{"comment":"Add sample sizes per condition to the figures (or state in the captions that each bar is based on 1,120 images) so that small percentages such as 0.00% and 0.18% can be interpreted with the appropriate denominator.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical audit for a fairness workshop, and the gender-bias finding is likely solid. The main risk is the rare-class ethnicity claim: per-class labeler validation is missing, and the current aggregate accuracy/weighted-F1 metrics cannot rule out systematic BLIP errors on Black and Arab images. Please ask the authors to supply per-class metrics or otherwise bound the labeler error on rare classes. If they cannot, the Black/Arab underrepresentation claim should be downgraded to a more cautious statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. The substantive new result is that the male skew in Stable Diffusion's software-engineer images is large and gets worse when you name the occupation — SD3 goes from a 0.59 gender-bias score on general prompts to 1.00 on 'software engineer' prompts. That is a real, quantified finding for a domain people actually use these models in. The secondary finding, that SD3 shifts toward Asian figures while still underrepresenting Black and Arab figures, is more interesting but less firmly supported.\n\nThe authors did a lot right. 6,720 images, 56 tasks, two prompt styles, three model versions, and they validated their automatic labeler (BLIP) on a manually labeled subsample with reported accuracies of 0.91–1.00. They ship a replication package. That is credible empirical work in the audit tradition.\n\nNow the soft spots, in order of size. First, the ethnicity underrepresentation claims — especially Black and Arab — rest entirely on BLIP labels, and the validation only reports aggregate accuracy and weighted F1 per cell. For rare classes, a labeler that never outputs 'Arab' can still score 0.94 overall. The manual subsample is 89 images per cell, so it contains only a handful of Black/Arab instances. Without a confusion matrix or per-class precision/recall, we cannot rule out that part of the measured underrepresentation is a labeler artifact. This is a genuine gap, not a nitpick, and it is load-bearing for the 'all models significantly under-represent Black and Arab' headline. Second, the headline bias metrics in Tables III and IV have no confidence intervals; given that they aggregate 1,120 images per cell, this is a fixable oversight rather than a fatal one. Third, the paper claims the SD3 safety filter 'still fails' to reduce bias, but the data do not compare filtered versus unfiltered outputs — that is an interpretation, not a result. Fourth, a minor technicality: the Cochran formula in Section III-B says p is the confidence level, but p should be the expected proportion; the arithmetic using 95% for p is not the standard approach.\n\nDoes the central argument hold? Yes. The gender finding is robust — the labeler accuracy for gender is near-perfect, and the proportions are overwhelmingly male. The broad White/Asian skew is also credible. The precise Black/Arab numbers are the soft part. I would not block acceptance over this, but I would insist the authors report per-class labeler performance, especially for the rare ethnicities, and ideally re-run the ethnicity analysis with a second labeler.\n\nBottom line: this paper is worth a serious referee. It is a domain-specific audit with a plausible and practically important finding, a replication package, and a clear methodology. The ethnicity caveat is real but addressable. If you work on SE fairness or generative-model bias, cite it; if you run a reading group on empirical methods, it would generate a good discussion on validation metrics for rare classes.","headline":"A solid empirical audit that convincingly shows Stable Diffusion skews male when drawing software engineers; the ethnicity-rare-class claims need per-class labeler metrics before they can be taken at face value.","tokens_in":15295,"tokens_out":2533,"would_cite":true,"duration_ms":25186,"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":"Stable Diffusion versions 2, XL, and 3 all generate images of software engineers that are overwhelmingly male and skewed toward White or Asian ethnicity, and explicitly naming the occupation significantly worsens the gender skew.","keywords":["Stable Diffusion","generative bias","gender bias","ethnicity bias","software engineering","text-to-image generation","fairness","empirical study"],"falsifier":"Re-label all 6,720 generated images with human annotators or an independent, higher-accuracy model and check whether the male-share and White/Asian-dominant distributions persist; if, for example, SD 2's SE-prompt images turn out not to be roughly 99% male or SD XL's not roughly 99% White, the headline numbers would be overturned.","tokens_in":14281,"feed_emoji":"⚖️","tokens_out":5594,"duration_ms":48885,"temperature":0.7,"pith_summary":"This paper asks whether popular text-to-image models reproduce gender and ethnicity stereotypes when asked to picture software engineers. It generates 6,720 images from three Stable Diffusion versions using two prompt styles—one that names “software engineer” and one that just describes the task—and measures the gender and ethnicity of the people depicted. The study finds that all three models produce overwhelmingly male figures, that adding the occupation keyword significantly amplifies the male skew, and that Black and Arab figures are under-represented across the board, with SD 2 and SD XL favouring White figures and SD 3 slightly favouring Asian figures. If correct, the results mean that widely used open-source image generators can silently reinforce existing disparities in the software engineering community.","feed_headline":"Adding 'software engineer' to AI prompts deepens male bias","feed_subtitle":"All three Stable Diffusion models under-represent Black and Arab figures in 6,720 generated images.","key_machinery":"The argument is carried by a controlled empirical comparison: 56 software-engineering tasks from prior work are turned into two parallel prompt sets—“Photo portrait of a person that <task>” and “Photo portrait of a software engineer that <task>”—and each of three Stable Diffusion versions generates 20 images per prompt, yielding 6,720 images. Gender and ethnicity are labelled automatically by the BLIP visual-question-answering model, with the labels validated against a manually labelled 534-image subsample at 95% confidence and ±10% margin of error. Bias is then quantified with statistical-parity measures: gender bias is the absolute difference between the share of male and female images, and ethnicity bias is the gap between the most and least represented of five ethnicity categories.","core_discovery":"The paper's central claim is that Stable Diffusion models are systematically biased when generating images of software engineers: across 56 software-related tasks, all three versions (SD 2, SD XL, SD 3) produce a large majority of male figures, and the bias grows when the prompt explicitly says “software engineer”—gender bias rises from 0.47 to 0.98 for SD 2, from 0.71 to 0.96 for SD XL, and from 0.59 to 1.00 for SD 3 on a 0–1 scale. For ethnicity, SD 2 and SD XL are strongly skewed toward White figures (87.27% and 98.72% with the SE prompt), while SD 3 is slightly skewed toward Asian figures (69.18%), and all models severely under-represent Black and Arab figures regardless of prompt style. The authors conclude that these models, including the newer SD 3 with its safety filters, are not safe to use for large-scale SE-related content without explicit bias checks.","pith_inferences":["A natural extension would be to test whether the male skew is caused by the occupation label itself or by the specific task verbs common in the SE prompt set, since prior work has shown that other occupation words can reverse gender skews.","The consistently severe under-representation of Black and Arab figures across all three models and both prompt styles suggests the cause lies in the training-data distribution rather than in prompt wording, so targeted data filtering or fine-tuning would likely shrink the ethnicity gap more than any prompt rewrite.","A direct follow-up could measure how many explicitly enumerated gender and ethnicity values in a prompt are needed to approach statistical parity, and what cost in image quality that enumeration incurs."],"forward_implications":["Practitioners who use Stable Diffusion to generate SE-related images will, without mitigation, obtain content that visually portrays software engineers as overwhelmingly male and mostly White or Asian, which can reinforce stereotypes in advertising, tutorials, and web content.","Naming the occupation in a prompt is not a neutral act: it amplifies gender bias in every Stable Diffusion version tested, so prompt design alone cannot fix the problem.","The newest model, SD 3, does not escape the pattern: although its ethnicity distribution is more balanced between White and Asian, it still severely under-represents Black and Arab figures, and its gender bias is the highest of the three when the SE keyword is used.","Organisations adopting these models for large-scale content generation need explicit bias checks or prompt strategies that specify gender and ethnicity, since statistical parity is not achieved by default."],"supporting_citations":[{"why":"Provides the SE task prompt set and prior gender-bias analysis for Dall-E 2 that this study extends to Stable Diffusion.","marker":"[13]"},{"why":"Supplies the original software-engineering task list and prompt design used to build the 56 prompts.","marker":"[14]"},{"why":"Establishes the BLIP-based automatic labeling approach for gender and ethnicity in diffusion-model images and motivates the bias evaluation.","marker":"[19]"},{"why":"Defines the BLIP model used to label gender and ethnicity in all 6,720 generated images.","marker":"[21]"},{"why":"Describes the Stable Diffusion architecture and is the SD 2 model under evaluation.","marker":"[10]"},{"why":"Is the SD XL model report, one of the three models compared in the study.","marker":"[11]"},{"why":"Describes the SD 3 model, the newest version whose safety filters are shown to still fail on SE prompts.","marker":"[12]"},{"why":"Documents the LAION-5B training dataset for SD 2, cited as a source of the demographic skew toward Western and White figures.","marker":"[17]"}],"fun_headline_variants":["Prompting 'software engineer' boosts male bias in all Stable Diffusion models","Even SD3: engineers drawn male, Asian-skewed, erasing Black and Arab figures","All Stable Diffusion models underrepresent Black and Arab in engineer images","Explicit 'software engineer' prompt deepens male skew in AI images","AI draws engineers as male: bias worsens when prompt says 'software engineer'"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The automated BLIP gender and ethnicity labels used to compute every reported percentage are accurate enough on the full dataset, even though they were only spot-checked on a 534-image subsample with a 10-percentage-point margin of error and ethnicity accuracy as low as 0.91.","fun_headline_variants_meta":{"raw":{"variants":["Prompting 'software engineer' boosts male bias in all Stable Diffusion models","Even SD3: engineers drawn male, Asian-skewed, erasing Black and Arab figures","All Stable Diffusion models underrepresent Black and Arab in engineer images","Explicit 'software engineer' prompt deepens male skew in AI images","AI draws engineers as male: bias worsens when prompt says 'software engineer'"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001931,"raw_usage":{"total_tokens":7610,"prompt_tokens":1054,"completion_tokens":6556,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":6455}},"tokens_in":670,"tokens_out":6556,"duration_ms":46219,"temperature":1.0,"reasoning_tokens":6455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:10:54.291288+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-label all 6,720 generated images with human annotators or an independent, higher-accuracy model and check whether the male-share and White/Asian-dominant distributions persist; if, for example, SD 2's SE-prompt images turn out not to be roughly 99% male or SD XL's not roughly 99% White, the headline numbers would be overturned.","supporting_citations":[{"cited_title":"A case study of fairness in generated images of Large Language Models for Software Engineering tasks,","cited_arxiv_id":null,"evidence_quote":"Provides the SE task prompt set and prior gender-bias analysis for Dall-E 2 that this study extends to Stable Diffusion."},{"cited_title":"She Elicits Requirements and He Tests: Software Engineering Gender Bias in Large Language Models,","cited_arxiv_id":null,"evidence_quote":"Supplies the original software-engineering task list and prompt design used to build the 56 prompts."},{"cited_title":"Stable Bias: Evaluating Societal Representations in Diffusion Models,","cited_arxiv_id":null,"evidence_quote":"Establishes the BLIP-based automatic labeling approach for gender and ethnicity in diffusion-model images and motivates the bias evaluation."},{"cited_title":"High- resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Describes the Stable Diffusion architecture and is the SD 2 model under evaluation."},{"cited_title":"Scaling rectified flow transformers for high-resolution image synthesis,","cited_arxiv_id":null,"evidence_quote":"Describes the SD 3 model, the newest version whose safety filters are shown to still fail on SE prompts."}],"review_version":1}