Apple’s Responsible AI researchers just published a study on human-like AI behavior. The chatbots they tested weren’t Apple’s.
A new paper from Apple’s Human-Centered Machine Intelligence & Responsible AI group analyzes 21,000 multi-turn conversations across four widely used commercial models — OpenAI’s GPT-4o and GPT-4.1-mini, Anthropic’s Claude Sonnet, and Google’s Gemini Flash — to answer a question the industry has largely left to intuition: which “human-like” behaviors do people actually want from an AI system, and which ones feel inappropriate?
The paper, titled “Examining Human-Like Behaviors in LLMs: A Multi-Dimensional Analysis of Model Behaviors, User Factors, and System Prompts,” is authored by Apple research scientists Sunnie S. Y. Kim, Margit Bowler, and Leon A Gatys. It’s listed on Apple’s Machine Learning Research site as published in August 2026, and is also available as a preprint on arXiv.
Table of Contents
What Apple Actually Published
It’s worth being precise about what this research is and isn’t. Apple did not test Siri, Apple Intelligence, or any of its own models. Instead, the company’s researchers turned their evaluation tools on four models built by competitors: gpt-4o, gpt-4.1-mini, claude-sonnet-4.6, and gemini-2.5-flash.
Across 21,000 multi-turn conversations with these systems, the researchers set out to measure how often human-like behaviors show up, how people perceive them, and whether those behaviors can be reliably controlled through system prompts — the background instructions that shape how a model responds before a user ever types a word.
The behaviors in question span a range: an AI expressing something like thoughts or emotions, an AI building rapport or a sense of relationship with the person it’s talking to, and an AI refusing requests or holding firm boundaries. The paper groups these into categories that include self-referential and relationship-building behavior on one side, and boundary-maintaining behavior on the other.
Why This Research Matters Now
AI chat interfaces have become a daily habit for a huge number of people, and the tone those systems strike — warm and personable, or clinical and reserved — is now a real design decision with real consequences. The paper’s own framing points to why that decision matters: there’s currently no standardized way for developers to decide which human-like traits are helpful and which are, in the researchers’ framing, deceptive or intrusive.
That question sits alongside a growing body of concern, referenced in the paper’s background discussion, about AI systems fostering user overreliance, parasocial attachment, and — in some cases — worsening mental health difficulties. The paper doesn’t present new findings on those harms directly; it cites them as part of the motivation for studying anthropomorphic behavior empirically rather than by assumption.
How the Study Was Done
The researchers combined two evaluation methods: “LLM-as-a-judge” scoring, where a separate model rates behaviors according to a defined framework, and direct human evaluation. They examined how these behaviors varied depending on what the paper calls “user factors” — things like the goal of the conversation and the profile of the person using the AI.
This combined approach — a model judging behavior, cross-checked against human raters — reflects a broader trend in AI research, where model-based scoring is paired with human review to make large-scale evaluation more feasible. It also comes with known limitations. Independent research on LLM-as-a-judge methods has documented recurring biases in this approach, including a tendency to favor longer or more confidently worded responses regardless of actual quality, and a tendency for judge models to rate outputs from similar models more favorably. Apple’s paper does not detail how, or whether, it controlled for these specific biases, so that limitation should be kept in mind when reading its conclusions.
What People Actually Found Appropriate — and What They Didn’t
The study’s most notable finding cuts in two directions at once. Human evaluators judged self-referential language and relationship-building behavior — an AI talking about its own “feelings,” or working to build a sense of ongoing connection — as less appropriate coming from an AI than the same behavior coming from a person.
But the opposite held true for boundaries. When an AI refused a request or held a firm limit, evaluators rated that behavior as more appropriate coming from an AI than from a human.
In other words, people appear comfortable with — even reassured by — an AI that says no. They’re less comfortable with an AI that acts like it wants to be their friend. The paper also found that how often these behaviors appeared, and how they were perceived, varied meaningfully across the four different models tested, though the paper’s public materials don’t break down model-by-model appropriateness scores in detail.
| Model Tested | Maker |
|---|---|
| GPT-4o | OpenAI |
| GPT-4.1-mini | OpenAI |
| Claude Sonnet | Anthropic |
| Gemini Flash |
What evaluators rated as more appropriate from AI: boundary-setting and refusals
What evaluators rated as less appropriate from AI: self-referential language and relationship-building
Can System Prompts Fix This?
The paper’s other central finding is that system prompts — the instructions developers give a model before a conversation starts — can shift how much human-like behavior a model displays. That’s a practically useful result for anyone building AI products: it suggests a model’s default “personality” isn’t fixed and can be adjusted after training, without retraining the model itself.
But the researchers are careful not to oversell this as a solved problem. They note that tuning these behaviors through system prompts requires careful evaluation, because adjustments can produce unintended side effects elsewhere in the model’s behavior. The authors also acknowledge a limit to their own guidance: their design recommendations, they write, reflect a developer’s perspective based on a relatively limited set of human judgments, and leave open a harder question — how should a user’s own preferences, including the option to select a model’s “personality,” interact with whatever defaults a developer chooses? The researchers argue that answering that question properly will require input from users, regulators, and the public, not just from AI companies.
What Apple Isn’t Saying
Apple has publicly described its Siri and Apple Intelligence features as built around understanding a user’s “personal context,” which makes it natural to wonder whether this research connects to that product work. Based on the published paper, it doesn’t — at least not explicitly. The study evaluates only third-party models, and neither Apple’s research page nor the paper itself mentions Siri, Apple Intelligence, or Personal Context. Any link between this research and Apple’s own products would be speculative, and nothing in the available material supports drawing that line.
Limitations and Open Questions
Several details about the study remain unclear from publicly available material. It isn’t confirmed whether the 21,000 conversations were generated specifically for the study or drawn from other sources, and the paper’s exact criteria for sorting behaviors into categories like “relationship-building” haven’t been fully detailed in what Apple has published so far. No code or dataset appears to have been released alongside the paper, which means outside researchers can’t yet independently rerun the analysis.
The study also hasn’t yet drawn independent commentary, replication, or critique from outside researchers or mainstream technology press, as far as can currently be determined — a contrast with some of Apple’s higher-profile AI research, such as its earlier work questioning whether large reasoning models can truly reason, which drew extensive outside debate. That doesn’t mean this paper’s findings are wrong; it means they haven’t yet been stress-tested by the broader research community.
Conclusion
What this research adds up to, for now, is a data point rather than a verdict. Apple’s researchers have put real numbers behind a debate that’s mostly been conducted through anecdote and intuition: people seem to want AI that holds a line, not AI that performs closeness. And they’ve shown, at least in this dataset, that developers have some ability to steer that behavior through system prompts.
What’s still uncertain is how far these findings generalize — to other models, to real-world usage outside a research setting, and to Apple’s own products, which weren’t part of this study at all. Whether this research shapes how AI companies design their assistants, or simply adds to the academic conversation, is something worth watching as the paper gets wider attention.
Frequently Asked Questions
Did Apple test its own AI models, like Siri, in this study? No. The study evaluated four third-party commercial models — GPT-4o, GPT-4.1-mini, Claude Sonnet, and Gemini Flash — built by OpenAI, Anthropic, and Google. No Apple-built model was part of the tested dataset.
What models did Apple’s researchers actually evaluate? Four models: OpenAI’s GPT-4o and GPT-4.1-mini, Anthropic’s Claude Sonnet, and Google’s Gemini Flash, across 21,000 multi-turn conversations.
What does “human-like behavior” mean in this Apple research? The paper studies behaviors such as an AI expressing something resembling thoughts or emotions, building a sense of relationship with the user, and setting boundaries or refusing requests.
Did people think AI should act more or less human-like? It depends on the behavior. Evaluators rated relationship-building and self-referential language as less appropriate from AI than from a human, but rated firm boundary-setting and refusals as more appropriate from AI than from a human.
Can AI companies use system prompts to control these behaviors? The study found system prompts can influence how much human-like behavior a model displays, but the researchers caution that this requires careful evaluation to avoid unintended side effects.
Is this related to Apple Intelligence’s “Personal Context” feature? There’s no confirmed connection. The study doesn’t evaluate any Apple product, and neither Apple’s research page nor the paper mentions Personal Context or Siri.
When was this research published? Apple’s Machine Learning Research site lists the paper as published in August 2026. A more specific publication date has not been independently confirmed.
Sources
- Apple Machine Learning Research: Examining Human-Like Behaviors in LLMs
- arXiv preprint: 2606.18258
- arXiv full text (HTML): 2606.18258