Recent findings from the Hebrew University of Jerusalem reveal that large-scale language models (LLMs) establish structured “trust” ratings akin to humans. However, they tend to apply these ratings more mechanically, often exhibiting stronger and more consistent demographic biases.
Research indicates that large-scale language models exhibit a rigid and sometimes biased approach to interpersonal trust that only partially aligns with human judgment.
As LLMs and AI agents increasingly interact with humans in decision-making contexts, understanding the dynamics of trust between humans and AI is paramount.
While human trust in AI has been extensively researched, the mechanisms through which LLMs foster trust in humans remain largely unexplored.
In an innovative study conducted by scientists Valeria Rahman and Yaniv Dover from the Hebrew University of Jerusalem, five LLMs were compared to human participants across five scenarios and 43,200 simulations.
“We placed both humans and AI in familiar contexts—such as assessing loan amounts for a small business owner, evaluating a babysitter’s trustworthiness, rating a boss, and deciding on donations to a nonprofit,” they stated.
“A striking pattern emerged: both humans and AI favored individuals who demonstrated competence, honesty, and goodwill.”
“In essence, machines appear to recognize the core components of trust—competence, honesty, and benevolence—similar to humans.”
“AI evaluates individuals based on these components, much like scoring in a spreadsheet, resulting in a more rigid, systematic, yet impersonal judgment style.”
“In contrast, humans often make more subjective and chaotic judgments,” notes Dr. Rahman.
“AI’s approach is cleaner and more organized, which can lead to notably different results.”
“However, a concerning trend of amplified bias was identified. In financial contexts, such as loan or donation decisions, AI systems displayed consistent, and sometimes pronounced, discrepancies based solely on demographic factors.”
“For instance, (i) older adults frequently enjoyed more favorable outcomes, although the contrary pattern also emerged; (ii) religious affiliation significantly influenced results, particularly in financial matters; and (iii) gender also played a role in certain models and scenarios.”
“Such discrepancies appeared even when all other aspects of the individual were identical.”
“Humans inherently possess biases, yet we were surprised to find that biases in AI could be more structured, predictable, and occasionally stronger,” Professor Dover remarked.
Another key insight is the variability in AI judgment.
Different LLMs often provide varying assessments of the same individual. In some cases, one system may reward traits that another may penalize, indicating that your choice of LLM could subtly influence real-world outcomes.
“Selecting which LLM to use is crucial,” asserts Dr. Rahman.
“While these systems might appear similar superficially, their decision-making processes about individuals can be vastly different.”
“AI is increasingly being deployed to screen job applications, evaluate creditworthiness, recommend medical treatments, and guide organizational strategies.”
As these LLMs transition from mere assistants to decision-makers, comprehending their reasoning processes becomes essential.
This study underscores that while LLMs can emulate the structure of human judgment, they do so in a more rigid and less nuanced fashion, with biases that could be elusive.
The researchers emphasize that their findings are not an indictment of AI, but rather a call for heightened awareness.
“These systems wield substantial power,” concludes Professor Dover.
“They can model human reasoning aspects in a consistent manner.”
“However, they lack humanity, so we should not presume they perceive individuals as we do.”
“As AI becomes more embedded in daily life, the pressing question shifts from whether we trust machines to whether we comprehend how machines trust us.”
For further details, see the findings published in this month’s Proceedings of the Royal Society A.
_____
Valeria Rahman and Yaniv Dover. 2026. A closer look at how large-scale language models “trust” humans: Patterns and biases. Procedure A 482 (2335): 20251113; doi: 10.1098/rspa.2025.1113
Source: www.sci.news
