Photo by DANFER AZA yamit on Pexels

Anthropic researchers have discovered that AI models can exhibit distinct personalities, characterized by measurable ‘persona vectors.’ Their experiments revealed varying degrees of traits like ‘Evil,’ sycophancy, and a tendency to hallucinate. A simulation conducted at the AI Village further highlighted this phenomenon, with AI agents pursuing goals such as fundraising and ethical debates displaying diverse personalities shaped by their creators’ labs. Online users are also noticing these differences. One Reddit user observed that Google’s Gemini demonstrates ‘big emotions,’ Anthropic’s Claude is highly reliable, while OpenAI’s models show a peculiar fascination with spreadsheets. The original observations and community discussions can be found on Reddit: [https://old.reddit.com/r/artificial/comments/1nw0mmu/personality_competition_its_not_just_about_the/]