Playing Bias: Epistemic Games for Interpreting Text-to-Image Models

Research from arXiv's Computers and Society domain introduces an epistemic framework to evaluate bias in text-to-image (T2I) artificial intelligence systems. The authors argue that while conventional technical audits effectively quantify demographic disparities, they fail to cultivate the critical literacy required to evaluate how generative outputs encode stereotypes and confer authority through visual staging such as pose, camera angles, props, costume, and settings. To address this, the study introduces a three-part humanities-informed framework: Interpret, Interrogate, and Expose, using structured prompt techniques to surface and debate cultural defaults.

This research matters because generative AI systems encode and circulate subtle representational defaults that cannot be resolved solely through statistical parity metrics. Organizations deploying generative visual media face reputational, cultural, and ethical exposure if they lack qualitative mechanisms to interrogate how models construct authority and visual norms. Developing qualitative auditing methodologies is critical for comprehensive AI risk management and governance.

Text-to-image systems establish normative visual defaults across elements such as pose, camera perspective, costume, props, and setting.

Forward consideration, not a verified fact.