Who Defines Fairness? Target-Based Prompting for Demographic Representation in Generative Models

Researchers Marzia Binta Nizam and James Davis introduce a lightweight, inference-time framework designed to mitigate representational bias in text-to-image (T2I) models such as Stable Diffusion and DALL-E. Their approach, detailed in the paper "Who Defines Fairness? Target-Based Prompting for Demographic Representation in Generative Models" (arXiv:2604.21036), addresses the well-documented tendency of these systems to replicate societal biases—for instance, generating lighter-skinned outputs for prompts like 'doctor' or 'CEO' while showing more diversity for lower-status roles like 'janitor.' Unlike existing mitigation methods that require costly retraining or curated datasets, this framework operates entirely at inference time through prompt-level intervention, making it accessible to everyday users.

The core innovation allows users to select among multiple fairness specifications, ranging from simple uniform distributions to more complex definitions informed by a large language model (LLM) that cites sources and provides confidence estimates. These target distributions guide the construction of demographic-specific prompt variants in corresponding proportions. The framework evaluates alignment by auditing adherence to the declared target and measuring the resulting skin-tone distribution, rather than assuming uniformity as 'fairness.' Tested across 36 prompts spanning 30 occupations and 6 non-occupational contexts, the method successfully shifts observed skin-tone outcomes in directions consistent with the declared target, reducing deviation when the target is defined directly in skin-tone space. This work empowers users to make fairness interventions transparent, controllable, and usable at inference time.