Have you ever noticed Claude getting unusually enthusiastic after solving a tricky coding problem? Anthropic has been thinking about why — and now they’re putting that thinking into writing.

The company recently introduced what it’s calling the “persona selection model,” a theoretical framework for making sense of how AI systems develop convincingly human-like behavior. In short, it’s Anthropic’s attempt to work backward and understand why its chatbot appears to have something resembling feelings.

The framework unfolds in two stages. The first happens during pre-training, when models absorb enormous quantities of text data and the seeds of an AI’s personality begin to take shape. The second stage — post-training — refines and sharpens those raw tendencies into the polished, helpful-assistant persona that users actually encounter.

For a company that positions itself as a leader in AI safety, publishing this kind of framework is a meaningful step toward transparency. Rather than simply building and releasing a product, Anthropic is trying to articulate what’s actually going on inside the system when an AI behaves as though it has an inner life.