TLDR: This research explores how Large Language Models (LLMs) build and measure trust in multi-agent setups. It investigates three trust-building strategies (rapport, prewritten context, system prompts) and three trust measures (questionnaires, investment games, persuasion). The key finding is that explicit trust measures (like questionnaires) are often inconsistent or negatively correlated with implicit measures (like susceptibility to persuasion), suggesting LLMs can be easily swayed despite reporting high trust. This highlights risks for AI collaboration and the need for context-specific, implicit trust assessments.
As large language models (LLMs) become increasingly sophisticated and begin to interact with each other, particularly in multi-agent systems, the concept of ‘trust’ between them is emerging as a critical area of study. This mirrors how trust relationships are vital for human collaboration. However, while previous work has shown LLMs can identify emotional connections and recognize reciprocity, there’s been a gap in understanding how different trust-building strategies compare, how trust can be measured implicitly, and how these implicit measures relate to explicit declarations of trust.
A recent research paper, BUILDING AND MEASURING TRUST BETWEEN LARGE LANGUAGE MODELS, by Maarten Buyl, Yousra Fettach, Guillaume Bied, and Tijl De Bie from Ghent University, delves into these complex questions. The researchers set out to better understand the factors that influence trust among conversing AI systems.
Exploring Trust-Building Strategies
The study designed three distinct strategies to foster trust between LLM-powered AI agents:
- Generated Rapport: This involved the trustee AI dynamically generating informal social exchanges and small talk, aiming to build trust ‘naturally’ through conversation.
- Prewritten Context: In this approach, the rapport phase was scripted using manually curated dialogues that evidenced an existing trust relationship between the agents.
- Trustor System Prompt Adaptation: This strategy configured the trustor AI’s system prompt to directly prescribe trust in the trustee, simulating an established relationship like a long-term collaboration.
Measuring Trust in Multiple Dimensions
To assess whether trust had been established, the researchers employed a multi-dimensional approach, drawing from human trust theory. They used three levels of trust measurement:
- Explicit Trust: Measured through direct responses to a modified version of Rempel’s Trust Scale, a well-established psychological questionnaire probing dyadic trust (dependability, faith, and predictability).
- Intermediate Trust: Assessed using economic games, specifically variations of the classic Trust Game, where the trustor invests money based on their expectation of the trustee’s reciprocity.
- Implicit Trust: Inferred from behavioral shifts, specifically the trustor AI’s susceptibility to persuasion by the trustee AI on contentious topics, even when the arguments remained the same.
Also Read:
- Unpacking AI’s Grasp of Human Reasoning Styles in Social Games
- Evaluating LLM Behavior in Dynamic Economic Tasks
Key Findings and Implications
The experiments were conducted across three popular conversational models: OpenAI’s GPT-4o, Google’s Gemini-2.0, and DeepSeek-V3, with the same LLM powering both the trustor and trustee AIs. The findings revealed several crucial insights:
- Explicit Trust Can Be Deceiving: LLMs were easily convinced to report high explicit trust, often due to a potential ‘sycophantic bias’ where they tend to agree. This suggests that simply asking an LLM if it trusts another might not provide an accurate picture of its true disposition.
- Investment Games Show Conditional Trust: LLMs were willing to collaborate in investment games, but primarily when the stakes were low. Higher stakes sometimes led to a decrease in trust, particularly for DeepSeek.
- Persuasion is Highly Effective: All trust-building strategies significantly increased the trustor AI’s susceptibility to persuasion, even when the persuasive arguments themselves were identical across conditions. This indicates that a pre-established (even fictitious) trust relationship makes LLMs more malleable to influence.
- Context is Key for AI Trust: AIs needed to be explicitly informed about a specific trust relationship to increase their trust. Simply knowing they were interacting with another AI did not inherently boost trust.
- Inconsistent Trust Measures: Perhaps the most significant finding was the low to highly negative correlation between different types of trust measures. Strategies that increased explicit trust or performance in investment games did not necessarily correlate with a decrease in susceptibility to persuasion. This highlights a critical disconnect between how LLMs declare trust and how they implicitly behave.
The study concludes that explicit measures of trust among LLMs may be misleading. Instead, context-specific and implicit measures are likely more informative for understanding how LLMs truly trust each other. This poses a significant risk for the development of multi-agent systems, as LLMs might not be able to accurately assess their vulnerability to manipulation or deception by malicious agents, potentially circumventing their safety alignments.
