Revealing Risk Preferences Through AI Prompting Effort

This paper analyzes “prompt engineering” through the economic lens of self-insurance against the risk of errors from noisy AI systems. To formalize this approach, we model an agent under cognitive load, allocating effort between working unassisted and prompting an AI assistant. The theoretical model demonstrates that an agent’s optimal prompting effort is driven by the agent’s attitude toward risk. Specifically, the model proves that risk-averse agents rationally “over-invest” in prompting effort, while risk-seeking agents “under-invest” relative to the risk-neutral benchmark. This outcome stems from the covariance between the marginal utility of performance and the marginal product of prompting. This alignment is positive for risk-averse agents, effectively boosting the AI’s perceived productivity. The novel implication is that prompting effort is an economically meaningful behavior that can be informative about an individual’s underlying attitude toward downside AI risk. These results offer a new perspective for understanding heterogeneity in AI adoption and oversight. They also suggest that, under comparable task conditions and controlling for prompting ability, observed prompting effort may be informative about attitudes toward downside AI risk. The framework therefore provides a risk-management perspective for understanding heterogeneity in AI governance in high-stakes settings such as healthcare and finance.