<p>Large language models (LLMs) are increasingly used as decision-support tools in domains characterized by risk and uncertainty, raising fundamental questions about how their behavior aligns with normative standards of rational choice. In this study, we develop and apply a structured battery of Expected Utility Theory (EUT)-based decision tasks to evaluate three widely used frontier models, ChatGPT, Claude, and DeepSeek, under both deterministic and non-deterministic settings. The battery spans five canonical domains of EUT violation, covering certainty effects, reflection effects in losses, mixed gambles, rare-event valuation, and insurance parity. Across models and conditions, we find that departures from expected utility are systematic, repeatable, and strongly model-specific, producing distinct and stable bias profiles that persist even when stochastic variation is removed. These findings extend the concept of bounded rationality beyond human cognition to collaborative intelligence systems in which decision outcomes emerge from the interaction of human users and AI models. Rather than acting as neutral correctives, LLMs instantiate distinct and persistent forms of bounded rationality and may, in some contexts, amplify existing cognitive distortions. We argue that these results have direct implications for bias management in human–AI decision-making systems. Structured EUT violation batteries provide a practical tool for diagnosing model-specific biases and should become a standard component of model evaluation and deployment in high-stakes settings. More broadly, our findings support a shift from attempts at bias elimination toward systematic bias management, combining normative diagnostics, model-specific testing, and sustained human oversight.</p>