Researchers Reveal Widespread Socioeconomic Bias in Leading AI Language Models

A new framework called SocioEval found that 13 frontier large language models exhibit substantial socioeconomic status bias in decision-making tasks, with bias rates ranging from 0.42% to 33.75% across different scenarios. The research reveals bias varies dramatically by context—lifestyle judgments show 10 times higher bias than education decisions—and current safeguards fail against domain-specific stereotypes. The findings underscore urgent need for class-based bias auditing in AI systems powering critical decisions.