Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2603.29317

Should I State or Should I Show? Aligning AI with Human Preferences

Keaton Ellis·Wanying Huang
View original source ↗Read source document ↗Export metadata ↓

Abstract

As AI agents become more autonomous, properly aligning their objectives with human preferences becomes increasingly important. We study how effectively an AI agent learns a human principal's preference in choice under risk via stated versus revealed preferences. We conduct an online experiment in which subjects state their preferences through written instructions ("prompts") and reveal them through choices in a series of binary lottery questions ("data"). We find that on average, an AI agent given revealed-preference data predicts subjects' choices more accurately than an AI agent given stated-preference prompts. Further analysis suggests that the gap is driven by subjects' difficulty in translating their own preferences into written instructions. When given a choice between which information source to give to an AI agent, a large portion of subjects fail to select the more informative one. Moreover, when predictions from the two sources conflict, we find that the AI agent aligns more frequently with the prompt, despite its lower accuracy. Overall, these results highlight the revealed preference approach as a powerful mechanism for communicating human preferences to AI agents, but its success depends on careful implementation.

Explore related subjects

econ.GNq-fin.EC

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Keaton Ellis, Wanying Huang. 2026-05-29. Should I State or Should I Show? Aligning AI with Human Preferences. https://arxiv.org/abs/2603.29317

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Local Media and the Shaping of Social Norms: Evidence from the Ebola outbreak

Media's influence on norms and behavior is widely recognized. Less is known about the role played by media being local. I examine this in a high-stakes context, the Ebola outbreak in Guinea. I exploit quasi-random variation in access to radio and the timing of a public-health campaign aired on community radio. I find that 12-17% of Ebola cases could have been prevented if places with access to a neighboring community radio station had instead had their own. Impacts are driven by radio being local rather than by ethno-linguistic belonging. Local media facilitates coordination in behaviors observed and sanctioned locally.

econ.GN↗
arXiv2026

Productivity Shocks and Input Misallocation: A Decomposition

This paper asks how much input misallocation productivity uncertainty generates and at which stage of input decisions it arises. I separate revenue productivity by when each component is revealed and trace each into the gap between an input's marginal revenue product and its price. In six European countries, shocks revealed after an input is committed account for 20 percent of capital gap dispersion and 5 percent of labor gap dispersion. An unanticipated one percent rise in productivity raises the capital gap by 0.92 percent and the labor gap by 0.19 percent, because most of the shock passes into the wage.

econ.GN↗
arXiv2026

When Do Type-Specific Wages Buffer Distributional Incidence in TANK?

When do relative wages buffer the unequal incidence of aggregate shocks? I derive a consumption-gap decomposition and a present-value condition for partial offset in a TANK model. An extension separates wage-setting demand elasticity from substitution between labor segments and allows each segment to contain both financial types. With a zero inherited wage gap and a same-sign discounted wedge, substitution above one gives offsetting earnings reallocation; substitution below one gives amplification. The channel disappears when financial types have identical segment exposure. Numerical experiments assess these mechanisms, shock persistence, policy feedback, and aggregate-IRF matching. In the nested perfect-alignment monetary benchmark, the peak consumption gap is about two-fifths smaller under type-specific wages than under the common-wage closure. These are conditional model comparisons, not empirical effect estimates or welfare rankings.

econ.GN↗