AI, Complexity, and Explanations
How to cite
Hümmecke, Eva; Andor, Mark; Damm, Simon; Feldhaus, Christoph; Fischer, Asja; Landes, Leon; Löschel, Andreas. 2025. "AI, Complexity, and Explanations." AEA RCT Registry. AEARCTR-0017444. https://doi.org/10.1257/rct.17444-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
As artificial intelligence (AI) systems increasingly make or influence decisions on behalf of individuals, understanding the conditions under which people accept and thus place trust in AI generated decisions becomes critically important. We conduct an incentivized experiment to causally estimate how task complexity and AI-generated explanations affect acceptance of AI interventions relative to human ones. Participants face lottery choices presented in a simple or a compound format. The lotteries are constructed to have different variances and expected values. An intervening agent, either a human (another participant) or an AI system, then selects the lottery that is played. After observing this intervened choice, participants decide whether to overrule it and choose themselves. We then elicit their willingness to accept compensation for keeping or overruling the agent’s choice. We further test whether explanations provided by the AI system influence acceptance of AI generated decisions. The design allows us to isolate the causal effects of complexity and explanations on acceptance for AI interventions relative to human decision makers.
Experimental design
Participants are randomly assigned to one of eight experimental groups in a 2×4 orthogonal design. The first dimension of the design varies the complexity of the choice participants face (low vs. high complexity). The second dimension varies the entity that intervenes in the decision (human intervention, AI intervention, explainable AI intervention (ex ante) and explainable AI intervention (ex post)). We then measure whether the humans or the AI’s decision is overruled. We also measure the WTA to keep the intervened decision.
Design details
Cited by
No citation data available for this registration yet.