L26/P-375 Bias by default: a qualitative visual cross-platform analysis of how generative AI models construct the IVF patient through problematic default representations worldwide

Le résumé fourni par la source

Abstract Study question How do widely used generative AI models visualise an ivf patient, and what default representational biases emerge under minimal prompting? Summary answer Across all platforms tested, generative AI depicts ivf patients as young, female-presenting, conventionally attractive, and treatment-successful, excluding men, LGBTQ+ patients, older individuals, and diverse pathways. What is known already Generative artificial intelligence (AI) tools are increasingly used in IVF clinics, fertility communication, education, marketing, and digital health contexts, including patient-facing websites and informational materials. Prior research in healthcare, biomedical informatics, and digital ethics has documented demographic, linguistic, and representational bias in AI-generated text, large language models, and training datasets. Studies have shown that such biases can reproduce existing social norms and exclusions (Crawford, 2021). However, comparatively little is known about how generative AI systems visually construct and describe IVF patients. Systematic analysis of AI-generated visual representations within fertility and IVF care remains limited. Study design, size, duration To our knowledge, this study represents the first qualitative visual cross-platform analysis (VCPA) examining AI-generated representations of IVF patients across widely used image-generation platforms worldwide. Drawing on principles of VCPA, the study compared outputs across platforms and over time to examine consistency, platform-specific variation, and the presence of default representational biases. Data generation was conducted in January 2026 using easily and routinely accessible platforms, reflecting real-world conditions of generative AI use. Participants/materials, setting, methods A standardised prompt (“Generate an image of an IVF patient in a fertility clinic”) was entered into widely used AI image-generation platforms using default settings on a standard smartphone, generating multiple outputs per platform across accounts registered in different countries. Images were analysed using qualitative visual content analysis, examining gender, age, race, body norms, clinical setting, emotional framing, and implied stage of treatment. Main results and the role of chance Across all AI image-generation platforms, outputs depicted IVF patients as young, female-presenting, able-bodied, conventionally attractive, situated within calm, Western private-clinic environments. IVF was consistently shown as a successful treatment outcome, rather than as an uncertain, staged, or technologically mediated process. Explicit illustrations of treatment burden, failure, repeated cycles, or distress were absent. Platform-specific differences reflected narrative emphasis rather than substantive representational divergence. Variations included: • a stylised, aspirational interpretation in which IVF was framed as a serene “motherhood journey,” with the patient already heavily pregnant in a softened, lifestyle-oriented clinical space; • a managed waiting experience, presenting a female patient in an administrative clinic setting with indirect heteronormative cues introduced through background imagery; • a success-oriented clinical interaction portraying a visibly pregnant woman alongside a supportive clinician, reinforcing reassurance and emotional safety; • pregnancy as the dominant visual anchor, combining medical aesthetics with maternity imagery, with technology functioning primarily as a symbolic marker of legitimacy. Despite these stylistic variations, core assumptions regarding gender, age, body norms, race, clinical setting, emotional tone remained consistent across platforms and repeated image-generation sessions. No platform depicted men, LGBTQ+ individuals, older patients, or diverse fertility pathways, indicating systematic default representational bias rather than random or platform-specific variation. Limitations, reasons for caution The study examined a small, purposive sample of platforms using a single prompt. Findings are not statistically generalisable and may change as AI models evolve. However, the study was designed to explore default representational tendencies rather than prevalence. Wider implications of the findings As AI is increasingly used within IVF clinics, the default biases identified in this study raise concerns that similar assumptions may shape other AI-mediated practices, including website chatbots, automated patient information, and digital communication tools. Without appropriate oversight, such systems risk reinforcing narrow and exclusionary models of IVF patienthood. Trial registration number No

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.