Trust is central to collaborative settings in which large language models (LLMs) are increasingly deployed. Yet little is known about whether LLMs exhibit a propensity to trust (PTT): a baseline tendency to extend or withhold trust that remains relatively stable across contexts. We investigate PTT in nineteen LLMs using two complementary approaches: a psychological self-report scale adapted from human research and a linguistic simulation framework designed to elicit trust-related decisions in context. While the questionnaire produces uniformly high PTT across models—likely reflecting social-alignment objectives and sycophantic response patterns—the simulation framework uncovers substantial, systematic differences in how models entrust others. Our simulations show that trust behavior is governed by the interaction between a baseline tendency to delegate and a model’s capacity to integrate cues about trustworthiness. More capable models, such as GPT-4o-mini, use such cues to adjust their decisions, allowing competence signals to modulate baseline tendencies. By contrast, other models, such as Llama-2-7B, exhibit stable delegation patterns that are largely insensitive to task-specific evidence, leading to systematic over-entrustment. These results show that performance depends not on baseline tendencies alone, but on how they are modulated by alignment-sensitive information. Ablation studies show that task-specific memory mechanisms enable models to better integrate trustworthiness cues, improving the calibration of delegation decisions. More generally, our findings show that questionnaire-based measures cannot disentangle baseline tendencies from context-sensitive adjustment, whereas behavioral simulations make this distinction observable.
Alice Plebe (Wed,) studied this question.