Research Reveals Large Language Models Exhibit Human-Like Trust Patterns and Biases

The findings highlight the potential for AI biases to affect critical decisions in finance, healthcare, and employment, impacting fairness and equity.

• Research published: A study by Valeria Lerman and Yaniv Dover analyzed how large language models (LLMs) assess human trustworthiness based on competence, benevolence, integrity, and demographic factors.
• Simulated experiments: The researchers conducted 43,200 simulations using five LLMs across various decision-making scenarios, revealing that LLMs exhibit amplified biases in trust assessments.
• Publication details: The study was initially posted on arXiv on April 22, 2025, and later published in the Proceedings of the Royal Society A on April 8, 2026.

• Trust dynamics: While human trust in AI has been extensively studied, AI's trust in humans remains underexplored, despite its growing role in high-stakes decisions.
• Bias amplification: The study found that LLMs tend to favor certain demographics, particularly Jewish, older, and male applicants in financial contexts, raising concerns about equity in AI applications.
• Model of trustworthiness: The research builds on Mayer et al.'s (1995) model, which includes competence, benevolence, and integrity, emphasizing the need for transparency in AI decision-making.

The research conducted by Lerman and Dover sheds light on the intricate dynamics of trust between large language models (LLMs) and humans. By employing a three-stage prompting protocol, the researchers elicited quantitative trust decisions from five different LLMs, including GPT-5 and Gemini 2.5 Pro, across five distinct scenarios. These scenarios ranged from evaluating a manager to assigning a babysitter, allowing for a comprehensive analysis of how LLMs assess human trustworthiness.

The study's findings reveal that LLMs predict trust with a high degree of accuracy, achieving an adjusted R² of up to 83.8% in the babysitter scenario. This level of predictive power indicates that LLMs apply a more modular and consistent judgment framework compared to humans, who often exhibit halo effects—where one positive trait influences the perception of other traits. Notably, LLMs placed a disproportionate emphasis on benevolence, suggesting that they may prioritize perceived good intentions over other critical factors like competence or integrity.

However, the research also highlights a significant concern: the amplification of demographic biases in LLM trust assessments. The models favored applicants based on age, gender, and religion, particularly in financial contexts. For instance, older individuals and those identified as Jewish were more likely to receive favorable evaluations. This bias raises ethical questions about the deployment of LLMs in sensitive areas such as finance and healthcare, where trust decisions can have profound implications for individuals and communities.

The implications of these findings are far-reaching. As LLMs become more integrated into decision-making processes, understanding their trust dynamics is essential for ensuring equitable outcomes. The study underscores the need for transparency in AI systems, as well as the importance of addressing biases that may arise from their training data and algorithms. As organizations increasingly rely on AI for critical decisions, the potential for biased outcomes necessitates a careful examination of how these systems assess human trustworthiness.

• Financial institutions: Risk assessment and loan approvals may be influenced by biased AI evaluations.
• Healthcare providers: Trust assessments in caregiving roles could affect patient outcomes and care quality.
• Employers: Recruitment processes may be skewed by LLM biases, impacting diversity and inclusion efforts.
• Policy makers: Regulatory frameworks may need to adapt to address AI bias in decision-making.

• Regulatory developments: Watch for new guidelines or regulations aimed at mitigating AI biases in decision-making processes, as these will shape how organizations deploy LLMs.
• Public discourse: Monitor discussions around AI ethics and bias, particularly in high-stakes sectors, as public sentiment can influence corporate practices and policies.
• Technological advancements: Keep an eye on improvements in LLM training methodologies that aim to reduce bias and enhance transparency in AI decision-making.