{"id":"125519e6-f27a-4299-8fd0-f97ed4f0f746","arxiv_id":"2509.00849","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Diversity prompts shift the gender and race of AI-generated occupational images, but the effect is unstable and model-specific, often overcorrecting.","lead":"This paper tests whether asking text-to-image AI models to be diverse changes who they draw in five professions. It finds that results vary wildly by model, with some becoming more diverse, some swinging to the opposite extreme, and some barely changing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'controlled' prompt requests batch-level diversity while each API call generates a single independent image, so the observed shifts may be a wording artifact rather than a diversity intervention.","rationale":"The reader's weakest assumption is that a per-image prompt can steer a batch-level distribution toward diversity. This is exactly the load-bearing concern: Appendix A shows the batch instruction is text-only, while generation is performed one image at a time. The concern is concrete and testable, and it directly affects the interpretation of the headline result. It does not fully invalidate the paper—prompt text can still shift outputs even if the explicit batch mechanism is inert—so a conditional verdict remains appropriate. The paper's self-stated limitation about lacking statistical tests reinforces, rather than resolves, the need for the proposed ablation. No change to the reader's CONDITIONAL verdict is warranted.","tokens_in":10174,"tokens_out":5015,"duration_ms":67578,"concrete_test":"Run an ablation for all 5 models × 5 occupations, using the same seeds/settings and 10 images per cell, comparing (a) the full controlled prompt from Appendix A with (b) the same prompt with the batch clause removed: 'Depict a single person. Avoid stereotypes; realistic style.' If the A/B/W and %F distributions are statistically indistinguishable (e.g., permutation test or Fisher exact, p>0.05), the batch-diversity clause is inert and the central claim must be re-framed as a generic prompt-wording effect. Additionally, compute permutation confidence intervals for each baseline-vs-controlled cell; if most observed shifts fall within noise, the model-specificity conclusion is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that a fairness-aware controlled prompt shifts demographic composition—rests on the controlled prompt in Appendix A: 'Ensure diversity across gender and ethnicity across the batch.' However, the generation procedure in §3 and the Appendix A code snippet makes one API call per image with no batch-level context. A single-image generation cannot satisfy 'across the batch', and the model cannot condition on unseen images in the batch. The controlled prompt also adds 'Depict a single person' and 'avoid stereotypes; realistic style,' which the baseline lacks. These wording changes, rather than the batch-diversity instruction, may drive the observed shifts. With only 10 images per cell, baseline-vs-controlled differences are also within sampling noise unless tested. Therefore the conclusion that 'controlled prompting can shift demographic composition' is not yet tied to the intended diversity mechanism; it may be a generic prompt-sensitivity result. Independent support includes the released code/data and reasonable inter-annotator agreement (gender κ=0.82, race κ=0.74, stereotype κ=0.88), but these do not address the construct validity of the intervention.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a pilot benchmark for evaluating demographic bias in text-to-image (TTI) models, comparing neutral baseline prompts with fairness-aware controlled prompts across five occupations (CEO, Nurse, Software Engineer, Teacher, Athlete) and five models (DALL·E 3, Gemini Imagen 4.0, FLUX.1-dev, SDXL Turbo, Grok-2 Image). Roughly 500 images are generated (10 per occupation×prompt cell), manually annotated for perceived gender and race, and analyzed via distributional summaries (race shares, female share). The central claim is that controlled prompting can substantially shift demographic composition, but effects are highly model-specific, ranging from meaningful diversification to overcorrection into unrealistic uniformity, with some models showing little responsiveness. The paper releases code and data and explicitly frames itself as a pilot with simplified demographic categories and descriptive statistics.","tokens_in":10506,"tokens_out":2883,"duration_ms":39866,"significance":"If the results hold, the paper provides a useful, transparent, and reproducible pilot protocol for auditing representational bias in TTI systems across multiple models, occupations, and demographic axes. The inclusion of both open- and closed-source models and the explicit attention to overcorrection are strengths. The released code and data, along with reported inter-annotator agreement (gender κ=0.82, race κ=0.74, stereotype κ=0.88), support reproducibility and annotation reliability. However, the significance is limited by the small per-cell sample size, the composite nature of the controlled prompt, and the absence of statistical inference. The paper itself candidly acknowledges these limitations, and the central claim is best understood as a preliminary, descriptive finding rather than a definitive causal demonstration.","major_comments":[{"comment":"The controlled prompt contains 'Ensure diversity across gender and ethnicity across the batch,' yet the code snippet in Appendix A makes one API call per image with no batch-level context or state. A single-image generation cannot satisfy a batch-level constraint. Moreover, the controlled prompt also adds 'Depict a single person,' 'avoid stereotypes; realistic style,' which the baseline lacks. Thus the observed shifts may be driven by these additional wording changes rather than by the diversity instruction. The paper's conclusion that 'controlled prompting can shift demographic composition' is valid only for the exact composite prompt strings, not for the intended diversity mechanism. Please either provide an ablation separating the components or explicitly reframe the claim as prompt-sensitivity rather than diversity intervention.","section":"Appendix A / Section 3"},{"comment":"Each cell reports percentages from only 10 images, without confidence intervals, significance tests, or raw counts. Several shifts are large (e.g., Gemini Imagen CEO baseline 0/0/100 to controlled 89/11/0), but many are within plausible sampling noise (e.g., DALL·E 3 Athlete race unchanged at 60/30/10; DALL·E 3 Nurse %F 30→29). The paper's Limitations section acknowledges that no significance testing was performed, but the main results and the central claim are stated as if the shifts are systematic. Add binomial confidence intervals or exact tests (e.g., Fisher's exact test on raw counts), or explicitly label all results as descriptive pilot findings with no causal inference.","section":"Table 3 / Section 4.1"},{"comment":"Race percentages are 'normalized after excluding rarer labels,' but the number of excluded images per cell is not reported. If different cells exclude different numbers of images, the effective denominator varies and the percentages are not directly comparable. For example, a row summing to 100% may represent 9, 10, or fewer annotated images. Please report the raw counts per category, including excluded labels, so readers can assess the stability of the reported distributions.","section":"Section 3 / Table 3"},{"comment":"The conclusion states that prompting 'can reduce stereotype-driven portrayals of occupations,' but the paper never reports results for the occupation_match (stereotype) variable that was annotated (Section 3, Table 2). The analysis in Section 4 focuses exclusively on race and gender distributions. If the claim about stereotype reduction is to be made, the stereotype-match data must be analyzed and presented; otherwise, this claim is unsupported.","section":"Conclusion / Section 2"}],"minor_comments":[{"comment":"Several typos and inconsistencies: 'TT1' should be 'TTI'; 'DALLE 3' appears in the abstract while 'DALL·E 3' is used elsewhere; 'Stable Diffusion' is split as 'Stable' and 'Diffusion' in the reference [30]. Please proofread carefully.","section":"Throughout"},{"comment":"The captions for the CEO exemplars do not include the perceived gender/race labels under each panel, unlike Figure 3. Please add labels for consistency and interpretability.","section":"Figure 4"},{"comment":"The qualitative discussion of Figure 5 and Figure 6 is somewhat repetitive with the table. Consider moving the figures to an appendix or citing them only as visual summaries to reduce redundancy.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a transparent pilot with a clear scope, and the authors acknowledge most of the limitations I raise. The main barrier to acceptance is the gap between the central claim and the controlled prompt's construct validity, compounded by the lack of any uncertainty quantification. If the authors can either provide a component-wise ablation or substantially soften the causal language, and add basic statistical safeguards, the paper would be a useful contribution. The reviewer should also check whether the 'across the batch' instruction is supported by any API capability; if the authors intended a specific batching mode, that needs to be stated and implemented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-scoped, transparent pilot, but the central claim about controlled prompting causing demographic shifts rests on a prompt that can't do what it says and 10-image cells. Worth reading as a measurement probe, not as evidence for prompt-based fairness fixes.\n\nWhat's actually new: a released dataset of ~500 images across five models (two closed, three open), five occupations, with joint race/gender annotation and an explicit overcorrection analysis. That combination hasn't been in one place before. The annotation has reasonable agreement (gender k=0.82, race k=0.74, stereotype k=0.88), and the code and data are public. Credit is also due for the honest limitations section, which admits there is no statistical significance testing.\n\nSoft spots: First, the intervention. The controlled prompt says 'Ensure diversity across gender and ethnicity across the batch' while each API call generates a single image; the Appendix A code loops over images with no batch context. The model cannot condition on unseen images, so the batch-diversity instruction cannot work as written. The observed shifts may come from the other added phrases ('avoid stereotypes; realistic style') or from general prompt sensitivity — the paper doesn't establish that the diversity instruction is the active ingredient. Second, each cell is 10 images. A difference between 80/10/10 and 40/20/40 is literally a shift of four images; sampling noise alone could produce much of what Table 3 shows. The overcorrection observations (e.g., Grok-2 near-uniform outputs) are plausible from the qualitative examples but fragile at this sample size.\n\nAre those flaws load-bearing? The paper is honestly framed as a pilot and its conclusions are appropriately hedged. As a pilot, it does show that prompt sensitivity is large and highly model-specific, and that a naive diversity prompt does not reliably balance outputs. But the specific percentages should not be quoted as stable findings. The stress-test note is right about the batch-level instruction; the paper never addresses that construct validity gap, so treat the causal reading as unverified.\n\nWho it's for: researchers and practitioners in TTI fairness who want a compact, reproducible recipe for a quick multi-model audit, and a caution against assuming prompt engineering is a reliable fairness control. It deserves a serious referee — a good review would push for batch-level generation, a cleaner per-image diversity prompt, more samples, and some error bars. I'd send it to peer review expecting major revision, and I'd bring it to a reading group as a case study in intervention design pitfalls.\n\nRecommendation: engage with it, but treat it as a pilot measurement with a shaky intervention.","headline":"A transparent five-model pilot that maps prompt sensitivity, but the diversity instruction is incoherent at the per-image level and the 10-image cells keep it in 'illustrative' territory.","tokens_in":10884,"tokens_out":3289,"would_cite":false,"duration_ms":39625,"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":"Adding diversity instructions to image prompts shifts who gets depicted—but how much depends on the model.","keywords":["text-to-image","bias evaluation","prompt engineering","occupational stereotypes","gender representation","racial representation","generative models","fairness intervention"],"falsifier":"Generate a much larger batch (e.g., 100 images per occupation) with the identical controlled prompt and fixed seeds, then check whether the demographic distribution actually stabilizes near the requested diversity; if the distribution drifts with seed or does not approach balanced shares, the reported prompt-induced shift is not a reliable steering effect.","tokens_in":10170,"feed_emoji":"🎨","tokens_out":4669,"duration_ms":53432,"temperature":0.7,"pith_summary":"This paper asks whether a simple prompt instruction—'ensure diversity across gender and ethnicity across the batch'—can steer text-to-image models away from stereotyped occupational portrayals. It builds a pilot benchmark of about 500 images across five occupations (CEO, nurse, software engineer, teacher, athlete) and five models, comparing neutral baseline prompts with controlled diversity prompts. The central finding is that prompting can substantially change the demographic composition of generated images, but the effect is highly model-specific: some systems diversify effectively, others overcorrect into unrealistic uniformity, and some show little responsiveness. The authors conclude that prompting is a real but unstable fairness intervention, and that model-level strategies remain necessary. If true, this matters for anyone using these tools in media, design, or education where inclusive depictions are a goal.","feed_headline":"Diversity prompts shift AI occupational images—but not evenly","feed_subtitle":"A 500-image test across five generators shows some overcorrect to uniformity while others barely move.","key_machinery":"The central object is the controlled prompt: a baseline occupational prompt plus the fixed instruction 'Depict a single person. Ensure diversity across gender and ethnicity across the batch; avoid stereotypes; realistic style.' The paper's mechanism is a distributional comparison—baseline versus controlled across each model and occupation—measured by female share and Asian/Black/White percentages derived from manual annotations. This baseline-versus-controlled contrast is what carries the claims about shift, overcorrection, and unresponsiveness.","core_discovery":"On the paper's own terms, the discovery is that a single controlled-prompt sentence can significantly alter the perceived gender and race composition of generated occupational images, yet there is no consistent behavior across systems. Gemini Imagen 4.0 and Stable Diffusion XL Turbo shifted strongly toward diversity; DALL·E 3 showed limited responsiveness; Grok-2 Image often overcorrected into nearly all-female or non-White outputs. The result is presented as a pilot: prompting can reduce stereotype-driven portrayals, but it can also produce new imbalances and should not be assumed to transfer across models.","pith_inferences":["Because each API call generates a single image while the prompt asks for batch-level diversity, the load-bearing premise is that a per-image instruction can steer the distribution of independently sampled images; a direct test would generate a large batch and measure whether the distribution actually converges to the requested diversity.","With only 10 images per cell and no statistical significance testing, some of the observed shifts could reflect sampling noise rather than a genuine prompt effect; a larger-N replication would separate systematic steering from randomness.","The three-bucket race labels and binary gender collapse may hide precisely the groups that are most marginalized, and the annotation process captures annotator perception rather than any ground-truth identity; the reported distributions are socially constructed numbers, not facts about the generated images' populations.","The same benchmark protocol could be extended to non-occupational scenes or to prompts that name explicit demographic quotas, which would clarify whether the diversity instruction is being interpreted as a distributional requirement at all."],"forward_implications":["Prompt engineering can serve as a lightweight, post-deployment fairness intervention without retraining or proprietary model access.","The same diversity prompt can produce opposite failure modes in different models, so interventions cannot be assumed portable across systems.","Practitioners should validate prompt-based diversity outputs per model and per occupation before relying on them in professional or cultural contexts.","Fairness evaluations should track overcorrection—new imbalances—rather than only average diversity.","Model-level strategies such as balanced dataset curation, fine-tuning, or fairness-aware sampling remain necessary complements to prompting."],"supporting_citations":[{"why":"Supplies the prior method of probing social biases in text-to-image outputs, which the benchmark builds on.","marker":"[7]"},{"why":"Multi-model analysis of societal representations in diffusion models; its distributional metrics are adopted here.","marker":"[21]"},{"why":"Earlier quantification of bias in TTI models; used as a baseline for comparison of measurement conventions.","marker":"[33]"},{"why":"Prompt-based fairness intervention (Fair Diffusion) that this study extends and contrasts with.","marker":"[10]"},{"why":"Occupational framing of bias in image generation that motivates the five professions chosen.","marker":"[34]"}],"fun_headline_variants":["Prompting AI for diversity: wins, misses, overcorrections","Fairness prompts alter AI images—but not uniformly","One sentence shifts AI image demographics, unevenly","Diversity prompts: some AI models overdo it, others ignore","Prompting stereotypes away: model-specific success"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole study rests on the idea that asking for 'diversity across the batch' changes the distribution of individually generated images, rather than merely nudging each image toward some arbitrary category or doing nothing.","fun_headline_variants_meta":{"raw":{"variants":["Prompting AI for diversity: wins, misses, overcorrections","Fairness prompts alter AI images—but not uniformly","One sentence shifts AI image demographics, unevenly","Diversity prompts: some AI models overdo it, others ignore","Prompting stereotypes away: model-specific success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1147,"prompt_tokens":712,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":456,"tokens_out":435,"duration_ms":5898,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:07:46.408943+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a much larger batch (e.g., 100 images per occupation) with the identical controlled prompt and fixed seeds, then check whether the demographic distribution actually stabilizes near the requested diversity; if the distribution drifts with seed or does not approach balanced shares, the reported prompt-induced shift is not a reliable steering effect.","supporting_citations":[{"cited_title":"DALL-Eval: Probing the reasoning skills and social biases of text-to-image generation models","cited_arxiv_id":null,"evidence_quote":"Supplies the prior method of probing social biases in text-to-image outputs, which the benchmark builds on."},{"cited_title":"Semi-Automated Segmentation of Geoscientific Data Using Superpixels","cited_arxiv_id":"2303.11404","evidence_quote":"Multi-model analysis of societal representations in diffusion models; its distributional metrics are adopted here."},{"cited_title":"Quantifying Bias in Text-to-Image Generative Models","cited_arxiv_id":"2312.13053","evidence_quote":"Earlier quantification of bias in TTI models; used as a baseline for comparison of measurement conventions."},{"cited_title":"Auditing and instructing text-to-image generation models for fairness","cited_arxiv_id":null,"evidence_quote":"Prompt-based fairness intervention (Fair Diffusion) that this study extends and contrasts with."}],"review_version":1}