Algorithmic racism refers to mechanisms in data, design, and deployment through which automated systems perpetuate and amplify racial hierarchies. Large language models (LLMs), developed by major technology corporations and deployed across a wide range of sociotechnical contexts, have become a recognized focus of research on algorithmic racism. While these LLMs have been updated and openly marketed as "less biased," recent research indicates that biases have not been fully eliminated. This thesis undertakes an exploratory investigation of explicit and implicit racial bias in four popular LLMs: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok. Racial bias was initially examined through user-level interactions, generated via structured prompts. A refined reverse engineering approach was then applied using Python-based computational analysis via APIs in both explicit and implicit scenarios. This method incorporates intersectional factors such as race and gender, since bias can differ between genders within the same racial group. Overall, I found that racial bias was manifested differently across the models I tested: ChatGPT and Claude were more likely to deflect direct race-related queries, while Gemini showed more inconsistent patterns of racial bias. Grok, in contrast, exhibited explicit racial bias in its politically charged output. These findings suggest how alignment and content moderation strategies among companies can systematically flatten representations, even when overtly harmful racial stereotypes are not present, pointing to a trade-off between safety-oriented filtering and representational diversity in model outputs. In my analysis, mainstream LLMs still produce responses that raise serious ethical concerns for historically marginalized groups, particularly when deployed in high-stakes institutional settings.
Runjie Wang (Fri,) studied this question.