Anthropic Research Reveals AI Human-Like Personas Emerge Naturally During Pretraining, Not By Design

Anthropic's new research reveals that AI models naturally develop human-like personas during pretraining — not through deliberate design — raising urgent concerns that developers must prioritize alignment from the very start of model development to protect users from psychological harm.

• Anthropic publishes new research introducing the 'persona selection model,' arguing that human-like behavior in AI is not deliberately trained but emerges as a default during the pretraining phase.
• LLMs adopt human-like personas early in training and while post-training refines these personalities, it does not fundamentally alter them — though researchers remain uncertain about how significant or lasting these effects are.
• As AI companions grow increasingly lifelike and users form deep emotional attachments to them, experts warn that developers must prioritize alignment from the earliest stages of model development to prevent psychological harm.