AAP 2026 CONFERENCE

I introduce an analytical framework for the fine-grained study of generative AI bias, applying the notion of social marking. When we say that a feature is socially marked, we mean that it stands out as unusual or noteworthy within a given social context, and that it prompts special treatment, for good or ill. I propose that much of the homogeneity of generative AI outputs results from the AI systems codifying that certain demographics and their features are marked, and in response to marked features AI systems will change the demographics it portrays, sometimes radically so. As a result, these systems will over-represent non-marked, dominant groups in the absence of certain features, and over-represent marked, non-dominant groups in the presence of such features. There are positive and negative forms of bias that result. The positive bias, where AI systems have stereotypical outputs, has attracted the most attention. But the negative bias is also noteworthy: some demographics are typically left out of depictions unless there is some marked feature in play, meaning that these systems selectively omit non-dominant groups in contexts where they should be visible.