Table 9_Evaluating diversity and stereotypes amongst AI generated representations of healthcare providers.xlsx

IntroductionGenerative artificial intelligence (AI) can simulate existing societal data, which led us to explore diversity and stereotypes among AI-generated representations of healthcare providers.
MethodsWe used DALL-E 3, a text-to-image generator, to generate 360 images from healthcare profession terms tagged with specific race and sex identifiers. These images were evaluated for sex and race diversity using consensus scoring. To explore stereotypes present in the images, we employed Google Vision to label objects, actions, and backgrounds in the images.
ResultsWe found modest levels of sex diversity (3.2) and race diversity (2.8) on a 5-point scale, where 5 indicates maximum diversity. These findings align with existing workforce statistics, suggesting that Generative AI reflects real-world diversity patterns. The analysis of Google Vision image labels revealed sex and race-linked stereotypes related to appearance, facial expressions, and attire.
DiscussionThis study is the first of its kind to provide a ML-based framework for quantifying diversity and biases amongst generated AI images of healthcare providers. These insights can guide policy decisions involving the use of Generative AI in healthcare workforce training and recruitment.

引言：生成式人工智能（Generative AI）能够模拟现有社会数据，由此我们得以探究AI生成的医疗从业者形象中存在的多样性与刻板印象。
方法：我们采用文本生成图像模型DALL-E 3，基于标注了特定种族与性别标识的医疗职业术语，生成了360张图像。随后通过共识评分法对这些图像的性别与种族多样性进行评估。为探究图像中存在的刻板印象，我们使用谷歌视觉（Google Vision）工具对图像内的物体、动作及背景进行标注。
结果：在满分为5分（5分代表多样性最高）的5分量表中，我们测得性别多样性得分为3.2，种族多样性得分为2.8，整体多样性水平适中。该结果与现有医疗行业从业人员统计数据相符，表明生成式人工智能（Generative AI）能够反映现实世界的多样性格局。对谷歌视觉（Google Vision）图像标签的分析显示，存在与性别、种族相关的刻板印象，涉及外貌、面部表情与着装。
讨论：本研究为首例提供基于机器学习（Machine Learning）的框架，用于量化AI生成的医疗从业者图像中的多样性与偏见的同类研究。上述研究结论可为涉及在医疗从业者培训与招聘环节使用生成式人工智能（Generative AI）的政策制定提供指导。