With the emergence and vertical integration of AI Agents into decision support systems and companies’ workflows, it is only a matter of time until these systems will be tasked with making partially or fully unsupervised decisions — from purchasing products and services to evaluating risks and outcomes of their own actions. For such autonomy to scale safely, we need to understand which fundamental biases large language models (LLMs) are susceptible to, and how these biases manifest in practice. Without this knowledge, organizations risk deploying agents that systematically misjudge probabilities, preferences, or interactions in ways that could amplify errors and create unforeseen consequences.
To address this challenge, we apply established methods from behavioral economics — traditionally used to study human decision-making — to the newest generations of LLMs. By designing and conducting behavioral experiments, we aim to uncover systematic biases not only in isolated model outputs but also in human/AI and AI/AI interactions. This includes studying how agents negotiate, cooperate, or compete, and how biases can cascade when multiple systems interact.
Our objective is to provide analyses and comparisons of LLM performance under decision-making conditions, identify key sources of bias, and propose mitigation strategies. Ultimately, this work will inform the design of more reliable and trustworthy AI agents capable of autonomous decision-making at scale.