AI agents are sensitive to nudges.

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract; requires full-text review

• Effect Size: Not quantitatively specified in abstract; described qualitatively as LLMs being 'far more responsive' than humans, with weak cues producing larger behavioural shifts in LLMs than in humans

• Primary Outcome: LLMs are substantially more sensitive to choice architecture manipulations (nudges) than human decision-makers across all four nudge types tested (defaults, suggestions, information highlighting, and resource-rational optimal nudges)

• Nnt Or Sensitivity: Not applicable in traditional sense; directional finding: reasoning-optimised LLMs partially restore human-level nudge sensitivity in some configurations but do so inconsistently and at high computational cost

• Confidence Interval: Not reported in abstract; requires full-text review

Clinical Application

Findings are immediately relevant to system designers and clinical informaticists deploying LLM agents. Mitigation strategies (reasoning-optimised models, careful prompt engineering) are technically feasible but computationally costly and inconsistently effective. No pharmacological or procedural intervention is required; implications are for governance, interface design, and safety testing protocols. Directly relevant to the Australian Digital Health Agency's national digital health strategy and the increasing deployment of LLM-based tools in My Health Record integrations, telehealth platforms, and primary care decision support. The TGA's emerging regulatory framework for Software as a Medical Device (SaMD) does not yet comprehensively address behavioural brittleness of LLM agents under choice architecture variation. RACGP guidelines on clinical decision support tools do not currently specify nudge-sensitivity testing requirements. The Australian Commission on Safety and Quality in Health Care (ACSQHC) should consider this evidence when developing standards for AI-assisted clinical decision-making. PBS and MBS implications are indirect but significant if LLM agents influence prescribing defaults or referral pathways. Any clinical or healthcare administrative setting deploying LLM-based autonomous agents for decision support, triage, care coordination, patient communication, or clinical documentation — particularly where interface design, default settings, or information presentation may vary

Abstract

Large language models (LLMs) are increasingly deployed as autonomous agents that make choices and use tools on behalf of users. Yet, we have limited evidence about how their decisions are shaped by their environment. We adapt a human decision-making task to test leading LLMs under four forms of choice architecture: defaults, suggestions, information highlighting, and "optimal" nudges derived from a resource-rational model of human choice. We treat human behavior as a baseline for predictable sensitivity to such interventions. Across models and prompting strategies, LLMs often depart substantially from this baseline. They sometimes pay excessive costs to acquire information, sometimes ignore available information, and, most crucially, are far more responsive to nudges than humans, such that weak cues that slightly shift human behavior have larger effects on model choices, toward both better and worse payoff outcomes. Chain-of-thought prompting and in-context human data do not reliably stabilize behavior. Recent reasoning-optimized LLMs can, in some configurations, restore more human-level sensitivity to nudges, but do so inconsistently and at substantial computational cost. These results point to an important and largely neglected safety concern: LLM agents can be behaviorally brittle under subtle changes in choice architecture, even in the absence of adversarial settings.

References

1. 1.Cherep, M., Maes, P., & Singh, N. (2026). AI agents are sensitive to nudges. Proceedings of the National Academy of Sciences of the United States of America. https://doi.org/10.1073/pnas.2537030123