What social stratifications in bias blind spot can tell us about implicit social bias in both LLMs and humans.

Large language models (LLMs) are the engines behind generative Artificial Intelligence (AI) applications, the most well-known being chatbots. As conversational agents, they-much like the humans on whose data they are trained-exhibit social bias. The nature of social bias is that it unfairly represents one group over another. Explicit bias has been observed in LLMs in the form of racial and gender bias, but we know little about less visible implicit forms of social bias. Highly prevalent in society, implicit human bias is hard to mitigate due to a well-replicated phenomenon known as introspection illusion. This describes a strong propensity to underestimate biases in oneself and overestimate them in others-a form of 'bias blind spot'. In a series of studies with LLMs and humans, we measured variance in perception of others inherent in bias blind spot in order to determine if it was socially stratified, comparing ratings between LLMs and humans. We also assessed whether human bias blind spot extended to perceptions of LLMs. Results revealed perception of others to be socially stratified, such that those with lower educational, financial, and social status were perceived more negatively. This biased 'perception' was significantly more evident in the ratings of LLMs than of humans. Furthermore, we found evidence that humans were more blind to the likelihood of this bias in AI than in themselves. We discuss these findings in relation to AI, and how pursuing its responsible deployment requires contending with our own-as well as our technology's-bias blind spots.