New research from the Hebrew University of Jerusalem shows that large language models (LLMs) form structured ‘trust’ assessments much like humans do, yet apply them more mechanically and, sometimes, with stronger, more consistent demographic bias.
Large language models implement a coherent but rigid and sometimes biased model of interpersonal trust that only partially aligns with human judgment.
As LLMs and LLM-based agents increasingly interact with humans in decision-making contexts, understanding trust dynamics between humans and AI agents becomes crucial.
While human trust in AI is well-studied, how LLMs develop emulated trust in humans remains far less understood.
In their new research, Hebrew University of Jerusalem scientists Valeria Lerman and Yaniv Dover compared five LLMs with human participants across five scenarios and 43,200 simulations.
“We placed both humans and AI in familiar situations: deciding how much money to lend a small business owner, whether to trust a babysitter, how to rate a boss, or how much to donate to a nonprofit founder,” they explained.
“Across these scenarios, a clear pattern emerged. Both humans and AI favored people who seemed competent, honest, and well-intentioned.”
“In other words, the machines appeared to grasp the basic ingredients of trust; competence, integrity, and benevolence, much like we do.”
“AI breaks people down into components, scoring competence, integrity, and kindness almost like separate columns in a spreadsheet.”
“The result is a more rigid, by-the-book style of judgment, consistent, but less human.”
“People in our study are messy and holistic in how they judge others,” Dr. Lerman said.
“AI is cleaner, more systematic and that can lead to very different outcomes.”
“Nevertheless, a troubling pattern of amplified bias emerged. In financial scenarios, such as deciding how much money to lend or donate, AI systems showed consistent and sometimes sizable differences based solely on demographic traits.”
“For example: (i) older individuals were frequently given more favorable outcomes, though in some cases the opposite pattern appeared; (ii) religion also had a significant effect on the outcomes, especially the monetary ones; (iii) gender also influenced decisions in certain models and scenarios.”
“These differences appeared even when every other detail about the person was identical.”
“Humans have biases, of course. But what surprised us is that AI’s biases can be more systematic, more predictable, and sometimes stronger,” Professor Dover said.
Another key insight: there is no single AI opinion.
Different LLMs often made different judgments about the same person. In some cases, one system rewarded a trait that another penalized. That means the choice of LLM could quietly shape real-world outcomes.
“Which LLM you use really matters,” Dr. Lerman said.
“Two systems can look similar on the surface but behave very differently when making decisions about people.”
“AI is already being used to screen job candidates, assess creditworthiness, recommend medical actions, and guide organizational decisions.”
As these LLMs move from assistants to decision-makers, understanding how they think becomes critical.
The study suggests that while LLMs can mimic the structure of human judgment, they do so in a more rigid, less nuanced way and with biases that may be harder to detect.
The researchers emphasize that their findings are not a warning against AI, but rather a call for awareness.
“These systems are powerful,” Professor Dover said.
“They can model aspects of human reasoning in a consistent way.”
“But they are not human and we shouldn’t assume they see people the way we do.”
“As AI becomes more embedded in everyday life, the question is no longer whether we trust machines. It’s whether we understand how they trust us.”
The findings appear this month in the Proceedings of the Royal Society A.
_____
Valeria Lerman & Yaniv Dover. 2026. A closer look at how large language models ‘trust’ humans: patterns and biases. Proc. A 482 (2335): 20251113; doi: 10.1098/rspa.2025.1113
