Biases in an artificial intelligence image-generator’s depictions of older adults, with or without dementia
Description
Background: This content analysis study investigates biases in image generation by two artificial intelligence (AI) tools, DALL-E 3 and Midjourney, specifically in portraying older adults and individuals living with dementia. Despite the widespread use of generative AI in various sectors, there is limited research on how these models might perpetuate stereotypes and stigmatization through their produced images. Methods: A total of 1,056 images were generated using specified prompts categorized into three groups: general older adults, dementia-related, and control. Each prompt began with "Photorealistic portrait" followed by specific scene descriptions. Two researchers conducted content analysis on each generated image, focusing on factors such as portrait style, setting, posture, apparent sex of subjects, and emotional affect. The analysis was executed with blinding and randomization protocols to ensure unbiased assessment. Chi-square tests assessed the relationship between prompt categories and the variables. Findings: Results indicated significant disparities in how older adults and those with dementia were depicted. DALL-E 3 generated more positive affect depictions of older adults compared to Midjourney. However, both models often portrayed dementia-related prompts with negative affect, suggesting a tendency to depict individuals with dementia in less favorable emotional states. Variations in depiction styles between the two AI models were noted, with DALL-E 3 showing a broader diversity of outputs. Interpretation: The findings highlight the potential of AI to reinforce stigmatizing stereotypes through biased image generation. Recommendations include carefully selecting prompts to avoid negative connotations and advocating for enhanced AI model transparency and inclusivity. Future research should expand to other AI models to develop a comprehensive understanding of biases and strategies for mitigation.