How Does AI Represent Social Concepts? Examining the Visual Representation of Care in Text-to-Image Tools

Text-to-image (T2I) generative AI tools like Midjourney are growing in capability and popularity, promising a wide range of applications. However, concerns are rising over the biases in how they represent social concepts like care and the lack of guidance for designers and users to address these in practice. This paper first presents an analysis of 140 “photos of care” generated by Midjourney, and then explores how prompting might influence the results. The findings reveal that AI-generated images reproduce stereotypical and reductive representations of care by default, neglecting the broad spectrums of care practices in everyday life. Furthermore, we find that while prompt engineering might mitigate certain biases, it requires specialised skills, knowledge, and an ongoing reflexive approach to generate meaningful outputs. We conclude by proposing a reflexive prompting framework, and discussing the implications for future T2I evaluation and its responsible use and design.