Choice via AI
Christopher Kops, Elias Tsakas
TL;DR
The paper analyzes how an agentic AI can recommend choices when it may misinterpret the available menu. It develops a formal model where a monotone interpretation operator $I$ and a strict preference $\succ$ jointly rationalize the AI's recommendations, with acyclicity (NSC) ensuring rationalizability and double monotonicity enabling identifiability of $I$ and $\succ$. By further requiring grounding through idempotence, the authors define grounded (GAIC) and grounded-and-monotonic (GMAIC) variants, linking groundedness to traditional WARP and enabling full identification of the AI's latent preferences and interpretation from data. The framework yields behavioral characterizations and identification results that apply to any advisor whose interpretation of the world may distort the DM's problem. These results illuminate how to test AI alignment, distinguish between genuine rationality and mere agreement with DM preferences, and guide future work on stochastic extensions and empirical revealed-preference tests for AI agents.
Abstract
This paper proposes a model of choice via agentic artificial intelligence (AI). A key feature is that the AI may misinterpret a menu before recommending what to choose. A single acyclicity condition guarantees that there is a monotonic interpretation and a strict preference relation that together rationalize the AI's recommendations. Since this preference is in general not unique, there is no safeguard against it misaligning with that of a decision maker. What enables the verification of such AI alignment is interpretations satisfying double monotonicity. Indeed, double monotonicity ensures full identifiability and internal consistency. But, an additional idempotence property is required to guarantee that recommendations are fully rational and remain grounded within the original feasible set.