According to Chris Olah on X (Twitter), citing Anthropic’s new research post, the persona selection model explains why AI assistants like Claude appear human by selecting consistent behavioral personas during inference rather than possessing subjective experience. According to Anthropic, the model predicts that large language models learn distributions over coherent social personas from training data and then condition on prompts and context to stabilize one persona, which yields human-like affect and self-descriptions without implying sentience. As reported by Anthropic, this framing clarifies safety and product design choices: steering prompts, system messages, and fine-tuning can reliably shape persona traits (e.g., cautious vs. creative), enabling controllability and brand-aligned tone at scale. According to Anthropic, measurable predictions include reduced persona drift under strong system prompts and improved user trust and satisfaction when personas are transparent and consistent, informing enterprise deployment guidelines for regulated sectors. As reported by Anthropic, this theory guides evaluation: teams can audit models with targeted prompts to surface undesirable personas and apply reinforcement or constitutional methods to constrain them, improving reliability, risk mitigation, and compliance in customer-facing workflows.