Apple study says LLMs' human-like behaviors need careful calibration
Apple's machine learning research team released a paper today that takes a hard look at how large language models mimic human behavior — and whether they should.
The study, published on Apple's ML Research blog, analyzed 21,000 multi-turn conversations across four major models: GPT-4o, GPT-4.1-mini, Claude Sonnet 4.6, and Gemini 2.5 Flash. The researchers examined a spectrum of human-like behaviors — from expressing emotions and building relationships with users to refusing requests and maintaining boundaries — and found these behaviors are pervasive across all models tested, though they vary significantly depending on the model and the nature of the conversation.
What makes this paper stand out is its focus on perceived appropriateness. Human evaluators in the study judged self-referential behaviors (models talking about themselves) and relationship-building behaviors (models forming bonds with users) as less appropriate when coming from an AI than when coming from a human. Boundary-maintaining behaviors — the refusals and guardrails that models use to decline requests — were actually rated as more appropriate from LLMs than from humans. The implication is clear: users don't want AI companions that act like people; they want systems that know their place.
The paper also explores controllability through system prompts, showing that developers can tune these behaviors — but with caveats. Adjusting one type of human-like behavior can have unintended effects on others, making it a balancing act rather than a simple switch. Apple's researchers provide a set of recommendations for responsible LLM design, pushing for more empirical grounding in how these systems are built rather than relying on intuition or user pressure to make models more "human."
This matters because the industry is rapidly pushing toward more conversational, emotionally aware AI — think companion apps, customer service bots, and digital assistants that feel less like tools and more like colleagues. Apple's research is a useful reality check: human-like doesn't always mean human-wanted. The paper suggests there's a meaningful gap between what models can do and what users actually find appropriate, and that closing the gap requires careful evaluation, not just more data.
What to watch: Apple has been conspicuously quiet about its own LLM ambitions compared to OpenAI, Google, and Anthropic. A research paper like this — measuring the social dynamics of competing models — signals Apple is paying close attention to the user experience layer even as it plays catch-up on raw capability.
Should AI labs be required to publish appropriateness benchmarks alongside model capability scores? Tell us in the comments.
