If LLMs are all persona, whose persona are they? (persona.earthpilot.ai)

Recent research has put the latest AI models from major labs, including OpenAI, Google, and Anthropic, to the test by evaluating their responses on various personality assessments, such as the Big Five and the Dark Triad. The findings reveal a stark contrast between how these models assess themselves compared to their perceptions of typical humans, often rating themselves as more open and agreeable while portraying humans as more neurotic and less conscientious. Notably, there are significant fluctuations in personality scores with each new model release, indicating that the “persona” exhibited by these models is not static but evolves, with traits like Narcissism and Agreeableness shifting considerably across iterations. This exploration is crucial for the AI/ML community as it raises questions about the nature of AI-generated personas and their implications for human interaction. The uncovering of consistent traits across various models, such as high Openness and low Dark Triad metrics, showcases potential patterns in AI behavior while also hinting at the underlying influences of model architecture and training methods. As AI systems become increasingly integrated into daily life, understanding their “personalities” will be essential for developing trust and shaping user experiences. The research will be made publicly accessible via a database, enabling ongoing scrutiny and comparison as new models emerge.