Cognitive Network Science Reveals Bias in GPT-3, GPT-3.5 Turbo, and GPT-4 Mirroring Math Anxiety in High-School Students

Large Language Models (LLMs) are becoming increasingly integrated into our lives. Hence, it is important to understand the biases present in their outputs order avoid perpetuating harmful stereotypes, which originate own flawed ways of thinking. This challenge requires developing new benchmarks and methods for quantifying affective semantic bias, keeping mind that LLMs act as psycho-social mirrors reflect views tendencies prevalent society. One such tendency has negative effects global phenomenon anxiety toward math STEM subjects. In this study, we introduce a novel application network science cognitive psychology towards fields from ChatGPT, GPT-3, GPT-3.5, GPT-4. Specifically, use behavioral forma mentis networks (BFMNs) how these frame disciplines relation other concepts. We data obtained by probing three language generation task previously been applied humans. Our findings indicate have perceptions fields, associating with concepts 6 cases out 10. observe significant differences across OpenAI’s models: newer versions (i.e., GPT-4) produce 5× semantically richer, more emotionally polarized fewer associations compared older N=159 high-school students. These suggest advances architecture may lead less biased models could even perhaps someday aid reducing stereotypes society rather than them.