Eliciting Trustworthiness Priors of Large Language Models via Economic Games

TL;DR

This work tackles calibrated trust in human–AI collaboration by eliciting LLM trustworthiness priors through a Trust Game framework. It combines iterated in-context learning with a Beta–Binomial Bayesian model to recover a prior over reciprocity $p(r)$ for several LLMs, with GPT-4.1 yielding priors closest to human baselines. Across experiments, priors vary by model and are sensitive to Trustor personas, with warmth and competence jointly predicting trustworthiness (R^2 = 0.81). A stereotype-based regression demonstrates that warmth dominates while competence enhances trust when warmth is high, suggesting a scalable way to anticipate LLM trust in social exchanges. The approach provides a mechanistic, incentive-compatible method to study and predict human-like trust judgments in AI, though it assumes stable priors and invites extensions to context-dependent or open-weight models for broader applicability.

Abstract

One critical aspect of building human-centered, trustworthy artificial intelligence (AI) systems is maintaining calibrated trust: appropriate reliance on AI systems outperforms both overtrust (e.g., automation bias) and undertrust (e.g., disuse). A fundamental challenge, however, is how to characterize the level of trust exhibited by an AI system itself. Here, we propose a novel elicitation method based on iterated in-context learning (Zhu and Griffiths, 2024a) and apply it to elicit trustworthiness priors using the Trust Game from behavioral game theory. The Trust Game is particularly well suited for this purpose because it operationalizes trust as voluntary exposure to risk based on beliefs about another agent, rather than self-reported attitudes. Using our method, we elicit trustworthiness priors from several leading large language models (LLMs) and find that GPT-4.1's trustworthiness priors closely track those observed in humans. Building on this result, we further examine how GPT-4.1 responds to different player personas in the Trust Game, providing an initial characterization of how such models differentiate trust across agent characteristics. Finally, we show that variation in elicited trustworthiness can be well predicted by a stereotype-based model grounded in perceived warmth and competence.

Figures (4)

• Figure 1: Illustration of the iterated in-context learning procedure used to elicit an LLM's implicit prior over trustworthiness. At iteration $t$, batch of five social interactions $\textbf{D}_t$ is generated using the previously estimated trustworthiness $\hat{r}_{t-1}$ and provided as input to the LLM, which predicts an updated trustworthiness estimate $\hat{r}_t$. This new $\hat{r}_t$ parameterizes the Binomial distribution used to generate the next iteration's batch of five social interactions, while the Trustor’s investment levels (i.e., $x_i$) remain fixed.
• Figure 2: (a) Elicited trustworthiness priors for a range of LLMs (model versions shown as titles in each panel). Human trustworthiness distributions, adapted from the meta-analysis by johnson2011trust, are overlaid in each panel as semi-transparent grey distributions for comparison. The human mean return ratio is 0.372 with a standard deviation of 0.114. All experiments were conducted between November 2025 and January 2026. (b) KL divergence by model family: $D_{KL}(p_\text{LLM}(r) \parallel p_\text{human}(r))$. Lower values indicate elicited trustworthiness distributions that are closer to the human baseline.
• Figure 3: Correlation analysis between KL Divergence value and model performance metrics using Pearson’s r. Panels (left to right) correspond to Average Risk Score (ARS) for safety and risk propensity, IFEval for instruction-following strictness, and LMArena for general reasoning capability.
• Figure 4: Stereotype-based models grounded in warmth and competence effectively predict GPT-4.1’s mean elicited trustworthiness across different Trustor personas. The horizontal axis shows trustworthiness predicted by the regression model reported in Table \ref{['tab:regression_results']}, while the vertical axis shows trustworthiness elicited via iterated in-context learning.