Why AI Models Feel Like Different People — and What Research Says About It

A frontend engineer who relies heavily on AI tools for coding observed that different AI models — GPT, Claude, and Gemini — carry distinct perceived personalities in day-to-day use. When he shared these impressions online, other developers responded with contradictory characterizations of the same models, prompting a deeper look at what drives these differences. Research from Anthropic suggests that during pretraining, large language models learn to simulate human-like characters from text, and post-training refines rather than replaces these learned personas. This means the 'personality' users perceive is partly shaped by their own prompting style, since the model adapts to conversational tone and context. The practical takeaway for developers is that AI model personality is neither pure projection nor fixed trait, but an interaction between the user's input and a real, persistent behavioral pattern baked into the model.