Synthetic faces generated by text-to-image (T2I) models are increasingly common in digital media, but how these models handle emotional prompts across demographic groups is not well understood. A new arXiv paper addresses this gap by systematically auditing emotion-driven demographic biases in synthetic face generation.
The authors examine how emotionally conditioned prompts—such as asking for a happy or grumpy expression—interact with age and gender. The paper's title, "Happy Young Women, Grumpy Old Men?", signals the kind of stereotype-laden outputs the audit is designed to uncover.
Because this is a single source, there are no conflicting findings to compare. The work contributes to a growing line of research on bias in generative AI, suggesting that even emotional descriptors can carry demographic assumptions that models amplify.