Researchers Reveal How AI Bots Judge Human Trustworthiness and Biases

A team led by Professor Yaniv Dover from the Hebrew University has investigated how AI language models assess human trustworthiness, competence, and kindness when making decisions such as granting loans or hiring babysitters. Their study found that bots evaluate these traits similarly to humans, with trustworthiness being the most critical factor. However, unlike humans, bots judge each trait independently and tend to give more extreme and consistent ratings.

The research also revealed that different AI models weigh these parameters differently, influenced by the cultural background and training data of their developers. For example, bots trained on American data may interpret directness as trustworthy, while others may see it as unpredictable. Demographic factors also affected AI judgments, with some biases detected such as favoring Jewish identity or showing gender-based preferences depending on the scenario.

Additional studies presented at recent conferences showed that AI can infer users’ ethnicity and gender from subtle cues and sometimes apply stereotypes in responses. Emotional warmth and perceived capability of users also influenced the quality of AI-generated replies. Furthermore, AI moral judgments align with human consensus but are susceptible to manipulation through changes in narrative perspective or social context.

These findings raise important questions about the growing role of AI in shaping human decisions and social interactions, especially as people increasingly rely on AI for moral and personal advice. While AI lacks genuine feelings or long-term memory of users, its judgments can still impact opportunities and relationships. Ongoing research aims to better understand and mitigate biases in AI evaluations to ensure fairer outcomes in an AI-integrated future.