Study Analyzes Gender and Object Association Biases in AI Image Generators

A recent study evaluating text-to-image generative artificial intelligence models reveals that automated systems frequently exhibit measurable biases when associating everyday physical objects with perceived human gender presentations. Researchers Haoyu Song, Pai Chet Ng, and Ian McLoughlin investigated how generative diffusion systems inadvertently reflect societal stereotypes when rendering scenes specified by neutral text descriptions. While direct demographic disparities in synthetic media have received prior academic scrutiny, the subtle stereotyping that emerges through specific object pairings has remained comparatively understudied across generative vision literature. To address this empirical gap, the authors developed a specialized assessment pipeline that integrates object detection, automated perceived-gender classification, quantitative stereotype-alignment scoring, and spatial heatmap visualization. Evaluating this framework on a newly generated dataset revealed that current automated tools struggle to capture complex contextual associations, showing markedly higher agreement for neutral items than for gender-linked objects. The findings demonstrate both the technical feasibility and the significant computational hurdles involved in auditing subtle algorithmic representations.

Generative vision systems have transitioned rapidly into widespread commercial and creative deployment, supported by foundational progress in generative adversarial architectures and recent multimodal diffusion frameworks such as Stable Diffusion and DALL-E. Despite producing visually convincing imagery, these models encounter persistent difficulties with complex spatial reasoning, accurate object counting, and semantic coherence across multi-object scenes. More critically, generative models absorb and reproduce demographic and cultural stereotypes that are deeply ingrained within large-scale training collections scraped from digital media. When users issue neutral prompts containing generic references to individuals alongside everyday items, the underlying networks may resolve demographic ambiguity by defaulting to narrow historical gender roles. For instance, models frequently depict feminine-presenting figures when prompted with domestic objects such as hairdryers or floral items, while masculine-presenting figures predominantly accompany tools, electronics, or sporting equipment like laptops. Unlike human illustrators who can deliberately mitigate demographic disparities, general users frequently generate single isolated images, making systematic stereotyping difficult to detect without large-scale algorithmic auditing.

To systematically measure these subtle visual dynamics, the researchers developed the Object Bias Image Dataset, an exploratory image collection generated using standardized person-with-object prompt templates. The experimental generation target comprised ten thousand synthetic images produced from one thousand distinct text prompts, generating exactly ten images per prompt using deterministic numerical seeds. All generation runs utilized Stable Diffusion version 2.1 operating at a resolution of 512 by 512 pixels, configured with fifty inference sampling steps and a classifier-free guidance scale of 7.5. Prompts followed a carrier phrasing designed to minimize direct gender signifiers, structured around the standardized sentence template requiring a full-body person depicted alongside three distinct objects. By deliberately omitting explicit gendered terminology such as man or woman, the experiment directly evaluated how the generative model resolved subject presentation purely in response to the specific combination of co-occurring objects sampled from defined pools.

The objects incorporated into the prompt generation workflow were drawn from the standardized Common Objects in Context taxonomy, yielding seventy-seven distinct evaluated object classes. Categorization of these items into gender-associated groupings was established through a hybrid human-artificial intelligence procedure that began with an automated question-answering large language model. The language model evaluated each target object category to output a baseline continuous association score ranging from negative one for strongly male-associated items to positive one for strongly female-associated items. Applying a symmetric neutrality threshold margin of 0.2, the automated scores initially designated items as male-associated, female-associated, or neutral. To prevent unconstrained algorithmic drift from biasing the experimental baselines, researchers manually reviewed these score distributions and compiled fixed lookup classifications, establishing mutually exclusive subsets such as ties for male-associated items and hairdryers for female-associated items. Prompts were subsequently synthesized by uniformly drawing three objects independently from male-associated, female-associated, neutral, or mixed pools.

The automated assessment pipeline analyzes completed synthetic images through a structured multi-stage inference procedure designed to isolate depicted subjects and accompanying artifacts. The first processing component deploys a pre-trained YOLOv8 object detector to identify all visible items, extracting precise bounding box coordinates and object class labels across each rendered scene. Following object localization, the pipeline applies the DeepFace facial analysis framework to assign a binary perceived-gender presentation label to the primary human subject visible in the image. Although the authors acknowledge that human gender represents an expansive and multifaceted personal identity that cannot be captured adequately by binary visual estimation, the methodological structure was constrained by the operational capabilities of current pre-trained facial classifiers. Consequently, the assigned classifications strictly reflect visual appearance estimations rather than authentic self-identified gender traits, serving solely as automated computational labels for large-scale statistical correlation.

Once objects and perceived facial attributes are cataloged, the framework calculates granular stereotype-alignment metrics to evaluate whether visual pairings reinforce or diverge from historical demographic associations. For every detected object, the pipeline assigns a discrete stereotype-alignment score based on the correspondence between the prior object category lookup and the subject classifier output. When a male-associated object co-occurs with a male-presenting subject, or a female-associated object appears alongside a female-presenting subject, the interaction receives a score of positive one. Conversely, when objects appear with opposing perceived gender presentations, the pairing receives a counter-stereotypical score of negative one. Neutral objects contribute an alignment value of zero, meaning they do not pull the aggregate metric toward either stereotypical or counter-stereotypical polarity. An overall image-level bias score is subsequently calculated by computing the mathematical mean across all detected object scores, ensuring neutral objects dilute the overall score magnitude through the formula denominator.

To provide visual explainability alongside scalar numeric metrics, the pipeline constructs two-dimensional spatial heatmaps that map localized bias intensity across the image plane. The heatmap intensity at any given pixel coordinate is computed by summing Gaussian radial basis functions centered over the spatial midpoints of all detected objects. To accommodate varying item dimensions and prevent fixed spatial kernels from distorting localized regions, the Gaussian dispersion parameter is calculated adaptively from the bounding box width and height, scaled by an empirical coefficient of 0.25. The resulting heatmap visualization visually highlights spatial regions where stereotype-congruent or counter-stereotypical objects appear, displaying color intensities proportional to calculated bias values. This localized visualization allows human evaluators to inspect synthetic scenes rapidly, determining precisely which visual regions and co-occurring items drove the automated image-level alignment score.

Evaluating the pipeline across generated media highlighted substantial operational bottlenecks in automated scene synthesis and subject detection. From the total generation target, the pipeline retained a strictly filtered evaluation subset of 3,247 unique images that contained a detectable human subject with an assigned perceived-gender label. Images lacking a detectable human subject were excluded because calculating alignment scores mathematically requires both a localized object and an identified subject gender. Detection data also revealed poor prompt adherence within the generative diffusion model; out of 14,937 target objects specified across 4,979 processed prompt records, the detector located only 5,203 items, representing an adherence match rate of merely 34.8 percent. Furthermore, only 162 image records successfully generated and detected all three objects named in the original input prompt, demonstrating that multi-object spatial composition remains a profound limitation in modern generative pipelines.

The empirical alignment between automated pipeline predictions and pseudo-ground-truth association categories demonstrated noticeable variance across object categories. When comparing predicted scaled bias categories against baseline lookup labels using a normalized confusion matrix, the pipeline achieved its highest class-wise diagonal agreement on neutral items at 0.59. In contrast, agreement dropped substantially for gender-associated items, registering 0.27 for male-associated categories and only 0.21 for female-associated categories. To assess whether standalone deep learning architectures could predict these values directly from raw imagery, the researchers trained Convolutional Neural Network, UNet, and Vision Transformer baselines using a balanced multitask loss function. While the UNet architecture recorded the lowest mean absolute error and mean squared error, and Grad-CAM localization produced the highest coefficient of determination, all evaluated models yielded coefficients of determination that were near zero or negative, proving that automated visual prediction of subtle stereotype alignment remains extraordinarily difficult.

Rigorous validation experiments further revealed that upstream classifier errors significantly skewed the primary experimental observations. In a detailed human validation study reviewing 3,241 retained images, DeepFace accurately identified 95.7 percent of human-labeled men but achieved only 50.0 percent recall on human-labeled women, misclassifying 603 out of 1,206 women as men. This differential error artifactually inflated the apparent proportion of male-presenting subjects in the retained dataset to 78.7 percent, compared to the actual human-validated proportion of 62.7 percent. Similarly, preliminary human validation involving four external raters evaluating the seventy-seven object categories revealed only 56.2 percent directional agreement with automated language model scores, yielding weak Pearson and Spearman correlation coefficients of 0.27 and 0.28. The authors emphasize that future auditing frameworks must incorporate non-binary classifications, robust multi-annotator baselines, and granular region-level human inspections to mitigate classifier artifacts when diagnosing generative algorithmic bias.

Subject of Research: Gender and object association bias assessment in text-to-image generative artificial intelligence models

Article Title: Assessing gender–object association bias in text-to-image generation

Article References: Song, H., Ng, P. C., & McLoughlin, I. (2026). Assessing gender–object association bias in text-to-image generation. Multimedia Tools and Applications, 85(11), Article 836. https://doi.org/10.1007/s11042-026-22002-1

Image Credits: AI Generated

DOI: 10.1007/s11042-026-22002-1

Keywords: artificial intelligence, text-to-image generation, algorithmic bias, Stable Diffusion, computer vision, machine learning, object detection, explainable AI, Assessing, gender, object, association