GPT Is a Nerd, Claude Is a Colleague: Why AI Models Have Personalities (and Why It Matters)

Short summary

A frontend engineer observes that GPT, Claude, and Gemini each feel like distinct personalities in daily use, and traces this to Anthropic's persona selection research: models learn human-like characters during pretraining, and post-training refines rather than replaces them. What we perceive as personality is partly our prompting reflected back and partly a real, steerable pattern inside the network. Understanding this helps practitioners choose models more deliberately.

• •Different AI models exhibit distinct 'personalities' shaped by both user prompting and underlying training mechanisms
• •Anthropic's persona selection research explains that models simulate human-like characters learned during pretraining, refined by post-training
• •Interpretability work like persona vectors shows traits correspond to identifiable, steerable directions inside the network

Generated with AI, which can make mistakes.