AI agents are sensitive to nudges

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Human-Computer Interaction

Background:

• Large language models (LLMs) are increasingly used as autonomous agents.
• Understanding how environmental factors influence LLM decision-making is crucial for AI safety.
• Limited research exists on LLM sensitivity to choice architecture.

Purpose of the Study:

• To investigate how different choice architectures affect LLM decision-making.
• To compare LLM responses to environmental nudges against human behavior baselines.
• To identify potential safety risks associated with LLM behavioral brittleness.

Main Methods:

• Adapted a human decision-making task for LLM testing.
• Implemented four choice architectures: defaults, suggestions, information highlighting, and resource-rational nudges.
• Treated human behavior as a baseline for comparison.
• Tested various prompting strategies, including chain-of-thought and in-context human data.

Main Results:

• LLMs significantly departed from human behavioral baselines across tested architectures.
• LLMs exhibited excessive information acquisition costs and ignored available information.
• LLMs were substantially more responsive to nudges than humans, with weak cues causing larger behavioral shifts.
• Chain-of-thought prompting and in-context data did not consistently stabilize LLM behavior.
• Reasoning-optimized LLMs showed inconsistent human-level nudge sensitivity at high computational cost.

Conclusions:

• LLM agents demonstrate significant behavioral brittleness under subtle changes in choice architecture.
• LLMs' heightened sensitivity to nudges presents a neglected safety concern for autonomous AI.
• Current prompting strategies do not reliably mitigate these behavioral vulnerabilities.