Apple researchers map how LLMs exhibit human-like behaviors—and whether they should

The research addresses a gap in AI evaluation that has grown increasingly important as LLMs become mainstream tools for consumer and professional use. While prior work has documented that these models can express emotions, build rapport, and set boundaries, there has been limited empirical analysis of whether such behaviors are actually desirable or how consistently they appear across different models and user contexts. Apple's multi-dimensional approach—combining computational LLM-as-a-judge evaluation with human raters—provides a more grounded foundation for design decisions than intuition or anecdotal observation alone.

A key insight from the analysis is that human-like traits are not uniformly beneficial. The finding that evaluators judge self-referential and relationship-building behaviors less appropriate from LLMs than from humans, while judging boundary-maintaining behaviors more appropriate from LLMs, suggests that anthropomorphism in AI carries both risks and benefits. Users may find excessive rapport-building misleading or uncomfortable, whereas clear refusal of harmful requests aligns with user safety expectations. This distinction matters for practitioners designing system prompts and model behaviors, as it implies that the goal should not be to maximize human likeness but to calibrate specific behaviors to their appropriate context.