AA-PersonalityBench measures AI models' personality traits
Summary
AA-PersonalityBench has identified distinct personality profiles among various AI models, highlighting three main clusters: OpenAI's GPT-6 models and Anthropic models, Google's models alongside Meta's Muse, and older OpenAI models with Llama 4 Maverick. The benchmarking method involved a panel of LLM judges evaluating 900 questions across 30 personality traits, which aggregate into eight indexes such as Expressiveness and Transparency. OpenAI's GPT-6 Astra and GPT-6.1 Sol, for instance, exhibit strong responsiveness in Transparency, while Anthropic models show less variation in their profiles. This analysis aims to enhance understanding of how these models communicate and interact with users.
Analysis
Meta: Meta develops AI models including the Muse family for research and product integration. The news highlights several Muse sub-clusters, with variants like Wonder and Spark showing strong or pronounced traits across most personality indexes compared to other labs. GPT-6: GPT-6 encompasses OpenAI's latest model generation with variants such as Astra and Sol. These show distinct clusters in the benchmark, with some versions ranking high on Transparency and others more average or lower across traits like Informativeness. OpenAI: OpenAI develops large language models including the GPT series used in various applications. In the news, its GPT-6 models form one emergent family with Anthropic models based on similar personality trait profiles, while older GPT versions cluster separately with Llama models. Anthropic: Anthropic creates the Claude family of AI models focused on advanced reasoning and alignment. Its measured models exhibit closely aligned personality profiles that are generally pronounced across indexes except Slop, with internal variation among Opus, Sonnet, and other variants. Muse Wonder: Muse Wonder is a high-performing model in Meta's Muse lineup. It appears in sub-clusters noted for strong representation on most personality indexes while being slightly lower on Transparency. Claude Opus 5.5: Claude Opus 5.5 is a top-tier model from Anthropic's Claude series. It ranks among the stronger Anthropic models across most personality dimensions in the benchmark results. Claude Fable 5.1: Claude Fable 5.1 is an Anthropic model evaluated in the personality benchmark. It shows the weakest profile among the listed Anthropic models across most indexes. Claude Sonnet 5.5: Claude Sonnet 5.5 is a core model in Anthropic's Claude lineup. The news positions it as intermediate among Anthropic models in overall personality profile strength. Model Clustering: Clustering reveals three main families: Anthropic models grouping with OpenAI GPT-6, Google models with Meta Muse, and older OpenAI models with Llama. Personality Indexes: The benchmark aggregates 30 traits into eight indexes including Expressiveness, Organization, Informativeness, Transparency, Warmth, Conviction, Imagination, and Slop. Benchmarking Approach: AA-PersonalityBench uses pairwise LLM judging on 900 questions across 30 traits to cluster models by communication style without system prompts.
Categories
machine_learningaitechai_agents