American == White in Multimodal Language-and-Image AI

AI models embed racial bias about American identity: CLIP, SLIP, and BLIP—three widely-used image-and-language AI systems—consistently associate "American" more strongly with White individuals than with Asian, Black, or Latina/o individuals, mirroring a bias documented in human psychology research.

The bias manifests across multiple AI tasks: When BLIP was asked to identify who is American, it labeled 97% of White individuals as American compared to only 3% of Asian individuals. For Asian individuals, the model incorrectly responded "China" as their state of residence 53% of the time, while always selecting American states for White individuals.

AI systems notice race selectively based on whiteness: In image captioning tasks, BLIP remarked on race 36% of the time when describing Asian individuals but never mentioned race when describing White individuals, suggesting the models treat White as a "default" or unmarked category.

The bias reflects problematic psychological associations: Beyond simple misidentification, the models associated White Americans with patriotism and being "born in America" more than other groups—and paradoxically associated White individuals with being less likely to treat people of all races equally, replicating harmful stereotypes from psychology research.

Standardized testing can expose hidden AI biases: Using controlled image sets from the Chicago Face Database and embedding association tests, researchers were able to systematically measure and document biases that might otherwise remain invisible in real-world AI deployments.