{"id":"017ef878-a575-4f2a-a1fd-d624430e58c4","arxiv_id":"2604.09923","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"GLEaN distills text-to-image model biases into single median-pixel portrait composites that communicate associations like skin tone and emotion as effectively as data tables but faster.","lead":"GLEaN generates many portraits from identity prompts in a text-to-image model, aligns them with facial landmarks, and creates a single median-pixel composite portrait that visually shows the model's typical output. A smart generalist might read it to see a practical way to make hidden AI biases understandable to non-experts without statistics or tables.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Median-pixel composition after landmark filtering may not faithfully capture the model's central bias tendency.","rationale":"The reader's weakest assumption directly targets the prerequisite for the headline user-study result. Because the study uses the GLEaN portraits as stimuli, any systematic distortion in those portraits undermines the claim that they communicate biases 'as effectively' as tables. The concern is internal to the method rather than external consensus; the paper's black-box reproducibility is a strength but does not address this representation fidelity issue.","tokens_in":1796,"tokens_out":313,"duration_ms":20169,"concrete_test":"For one prompt (e.g., 'a doctor'), generate the full set of images, compute both the reported GLEaN composite and an unfiltered median (or mean) over all raw outputs; have blinded raters score perceived skin tone, age, and emotion on both versions. If ratings differ by >15% on any attribute, the filtering step materially changes the represented bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The user-study claim requires that GLEaN portraits accurately convey the model's encoded biases. The pipeline's final step (median pixel after landmark-based filtering and alignment) can be distorted by common T2I artifacts (e.g., warped faces, inconsistent lighting), prompt sensitivity, or the exact filtering thresholds. If these distort the composite away from the true distribution of generated outputs, the study compares viewer response to an artifactual image rather than to the model's actual central tendency.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces GLEaN, a three-stage pipeline (large-scale generation from identity prompts, facial landmark filtering/alignment, and median-pixel composition) to produce representative portraits that visualize biases encoded in text-to-image models for non-technical audiences. Demonstrated on Stable Diffusion XL across 40 social/occupational prompts, the composites reproduce documented biases and surface new skin-tone/emotion associations. A between-subjects user study (N=291) claims the portraits communicate biases as effectively as data tables while requiring significantly less viewing time. The method is presented as model-agnostic and replicable on black-box systems.","tokens_in":1889,"tokens_out":581,"duration_ms":45343,"significance":"If the user-study results and the fidelity of the median composites hold, GLEaN would provide a scalable, accessible tool for public comprehension of T2I biases, filling a gap between technical auditing methods and lay audiences. Strengths include its black-box applicability, open-source release, and potential to support broader discourse on AI fairness without requiring statistical expertise.","major_comments":[{"comment":"Abstract and User Study section: The central claim that GLEaN portraits 'communicate biases as effectively as conventional data tables' with significantly less viewing time rests on a between-subjects study (N=291) but reports no details on design, controls, randomization, statistical tests, effect sizes, or power analysis. This omission prevents evaluation of whether the equivalence conclusion is supported.","section":"Abstract / User Study"},{"comment":"§3 (GLEaN Pipeline, median-pixel composition step): The assumption that the final median-pixel portrait after landmark filtering and alignment accurately represents the model's central tendency is load-bearing for all downstream claims, yet no validation is provided against the full distribution of generated images. Common T2I artifacts (warped faces, inconsistent lighting) or prompt sensitivity could systematically distort the composite away from the true bias distribution.","section":"§3 (GLEaN Pipeline)"}],"minor_comments":[{"comment":"The manuscript would benefit from explicit discussion of limitations, including sensitivity of results to the specific landmark detector, alignment method, and number of samples used in the median computation.","section":"Discussion / Limitations"},{"comment":"Figure captions and the GitHub link should include version numbers or commit hashes to support reproducibility of the exact pipeline used for the reported composites.","section":"Figures / Code Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper's emphasis on public legibility is a good fit for an interdisciplinary AI-ethics or HCI venue, but the missing statistical details on the user study raise concerns about whether the central effectiveness claim can be assessed without substantial additional material."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive and detailed review. We address each major comment below and have revised the manuscript to incorporate additional details and validation as requested.","responses":[{"response":"We agree that the original manuscript did not provide sufficient detail on the user study to allow full evaluation of the claims. In the revised version, we have expanded the User Study section to include a complete account of the between-subjects design, randomization procedures for condition assignment and prompt order, controls for attention and demographics, the statistical tests employed (independent-samples t-tests for viewing time and two one-sided tests for equivalence on comprehension accuracy), effect sizes, and a post-hoc power analysis. We also report participant exclusion criteria and make the anonymized data and analysis scripts available. These additions directly support evaluation of the reported findings.","revision_made":"yes","referee_comment":"[Abstract / User Study] Abstract and User Study section: The central claim that GLEaN portraits 'communicate biases as effectively as conventional data tables' with significantly less viewing time rests on a between-subjects study (N=291) but reports no details on design, controls, randomization, statistical tests, effect sizes, or power analysis. This omission prevents evaluation of whether the equivalence conclusion is supported."},{"response":"The referee correctly identifies that explicit validation of the median composite against the full image distribution was absent. While the median pixel operation is robust by construction, we have added a new validation subsection in §3 that compares each median portrait to the corresponding mean image and to a random sample of 50 individual generations using pixel-wise variance maps and perceptual similarity metrics (SSIM and LPIPS). The results show that landmark filtering substantially reduces artifact-induced variance and that the median better approximates the central tendency than the mean. We also discuss remaining limitations due to prompt sensitivity. This revision provides the requested empirical grounding.","revision_made":"yes","referee_comment":"[§3 (GLEaN Pipeline)] §3 (GLEaN Pipeline, median-pixel composition step): The assumption that the final median-pixel portrait after landmark filtering and alignment accurately represents the model's central tendency is load-bearing for all downstream claims, yet no validation is provided against the full distribution of generated images. Common T2I artifacts (warped faces, inconsistent lighting) or prompt sensitivity could systematically distort the composite away from the true bias distribution."}],"tokens_in":1504,"tokens_out":511,"duration_ms":32194,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a three-stage pipeline that generates many images from identity prompts, aligns them with facial landmarks, and collapses them into a single median-pixel portrait. This produces something a lay viewer can read at a glance, like a composite face for 'doctor' versus 'felon.' That presentation trick is new enough to be worth noting, even though it rests on standard generation and image-processing steps. The authors show it on Stable Diffusion XL for 40 prompts, recover known biases, and surface a skin-tone/emotion link. They also release the code, which is useful for anyone who wants to try it on closed models. The between-subjects study with 291 participants is the main empirical claim: the portraits match tables for bias communication but take less time. If that holds, it is a practical step toward public legibility of generative bias. The soft spots sit in the validation. The abstract gives almost no study design, controls, or statistical detail, so it is hard to judge whether the equivalence result is robust or whether demand characteristics played a role. More importantly, the median-pixel step after landmark filtering can be pulled off course by common generation artifacts, inconsistent lighting, or prompt sensitivity. If the composite ends up showing an artifact average rather than the model's actual central tendency, then the user study is comparing responses to a distorted image. That concern is not fatal, but it needs direct checks, such as comparing composites to raw distributions or testing sensitivity to filtering thresholds. The work is aimed at people who build or audit generative systems and want tools that non-technical audiences can use without reading tables. It is not a deep theoretical advance, but the visualization idea is straightforward and the black-box compatibility is a real plus. I would send it to review so the study protocol and artifact checks can be examined properly.","headline":"GLEaN gives a clean visual shortcut for showing T2I biases to non-experts, but the user-study evidence is still thin on methods and the median composites may not be as faithful as claimed.","tokens_in":2370,"tokens_out":448,"would_cite":false,"duration_ms":25535,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A pipeline that generates many portraits from a prompt, aligns them, and averages their pixels creates single images that let non-experts see biases in text-to-image models as clearly as data tables do.","keywords":["text-to-image","bias detection","explainability","generative AI","user study","median composition","facial landmarks","public understanding"],"falsifier":"A direct comparison where the GLEaN composite differs substantially from the most frequent visual features observed across the full set of generated images, or where participants detect biases less accurately with the portraits than with equivalent data tables.","tokens_in":2688,"feed_emoji":"🖼️","tokens_out":674,"duration_ms":36806,"temperature":0.7,"pith_summary":"The authors present GLEaN, a three-stage process for turning text-to-image model outputs into visual bias summaries. Large sets of images are created from prompts involving social or occupational identities, faces are aligned using landmarks, and a median composition is formed to represent the typical output. This visual format requires no statistical training to interpret, allowing viewers to directly observe associations the model makes. A controlled study with hundreds of participants confirmed that these portraits communicate bias information with the same accuracy as numerical tables while requiring less time to review. The technique depends only on generated images, enabling its use with any text-to-image system regardless of whether its internal details are accessible.","feed_headline":"Single portraits expose text-to-image biases without stats","feed_subtitle":"Averaging aligned faces from many generations matches data tables for bias detection but takes significantly less viewing time.","key_machinery":"The median-pixel composition step, which follows landmark-based filtering and spatial alignment of large-scale generated faces to create a single representative portrait of the model's typical output for a given prompt.","core_discovery":"GLEaN distills the central tendency of a text-to-image model's responses to identity prompts into median-pixel portraits by generating images at scale, applying facial landmark filtering for alignment, and composing the median across pixels. These portraits reproduce established biases and highlight additional patterns, such as links between skin tone and expressed emotion, and a user study demonstrates that they convey bias information to the public as effectively as conventional tables but in significantly shorter viewing times.","pith_inferences":["Extending the alignment and composition process to other image domains, such as objects or scenes, could broaden bias auditing beyond faces.","Integration into public dashboards might allow ongoing monitoring of model updates for bias shifts.","Comparing composites across different models could provide a standardized visual benchmark for bias levels."],"forward_implications":["Biases in model outputs become immediately visible to people without technical expertise.","The method can be applied to proprietary or black-box models using only their generated images.","New bias patterns can be discovered through visual inspection of the composites.","Communication of model behaviors to the public can be streamlined compared to tabular data presentations."],"fun_headline_variants":["Median portraits expose T2I biases without stats","Face composites detect text-to-image model stereotypes","Aligned portraits reveal prompt based biases visually","Portraits match tables for faster T2I bias detection"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The median pixel values after landmark alignment accurately capture the model's intended central representation rather than being altered by generation noise, prompt wording, or the specific alignment rules chosen.","fun_headline_variants_meta":{"raw":{"variants":["Median portraits expose T2I biases without stats","Face composites detect text-to-image model stereotypes","Aligned portraits reveal prompt based biases visually","Portraits match tables for faster T2I bias detection"]},"model":"grok-4.3","cost_usd":0.007081,"raw_usage":{"total_tokens":3318,"prompt_tokens":756,"num_sources_used":0,"completion_tokens":50,"cost_in_usd_ticks":70812000,"prompt_tokens_details":{"text_tokens":756,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2512,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":756,"tokens_out":50,"duration_ms":22087,"temperature":1.0,"reasoning_tokens":2512,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-10T16:42:36.214651+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A direct comparison where the GLEaN composite differs substantially from the most frequent visual features observed across the full set of generated images, or where participants detect biases less accurately with the portraits than with equivalent data tables.","supporting_citations":[],"review_version":1}