Large language models (LLMs) from major tech firms like OpenAI, Meta, and Google show covert racial biases against African Americans, especially in interpreting speech patterns.

A technique known as Matched Guise Probing revealed that LLMs tend to associate African American dialect with lower-prestige jobs, criminal convictions, and harsher sentences, irrespective of explicit racial identifiers.

Attempts to correct these biases through safety training and human feedback have proven to be ineffective or counterproductive.

Similar studies have identified racial bias in AI, such as OpenAI's GPT 3.5, which showed prejudice against African American names in a hiring context.

The persistent racial bias in AI models underscores an urgent need for more effective research and solutions to prevent potential harm, especially in business and legal settings.