Mehdi LOUAFI

This paper studies whether profile-based explanations influence investors' acceptance of algorithmic risk recommendations in a randomized controlled trial embedded directly in the platform's interface of a leading French robo-advisor. Users were assigned either to see graphical explanations of the drivers underlying their recommended risk score and associated portfolio or to receive the standard interface with no explanation. Our results, obtained in a real-world setting with actual clients of a FinTech, do not support the adherence gains from increased transparency that are widely anticipated in the literature. Overall, providing profile-based explanations is not found to increase acceptance of the recommended profile nor raise users' engagement with the platform. However, we find a heterogeneous treatment effect as profilebased explanations lead to a greater downward deviation among desktop users who have already deviated to safer-than-recommended portfolios, but this pattern disappears once users' experience of the platform is taken into account. We observe non-causal evidence in both conditions that behavior is shaped primarily by the digital context and experience: phone and first-time users are more likely to accept the portfolio recommendation than desktop and returning users. While such transparency-enhancing profile-based explanations are informative, they are not a universal lever for adherence, suggesting that explanation design should be tested and tailored across device types and users' experience.

The Digital Services Act and the AI Act adopted by European institutions require algorithmic decisionmaking systems to meet transparency obligations through the provision of explanations of their functioning. As algorithmic decision-making systems increasingly shape individuals' economic and social lives, this paper experimentally tests whether adding a non-technical explanation to a neutral system description affects public acceptance. The study relies on a large-scale survey experiment on nationally representative adult samples in France, Germany, and Italy in which each respondent evaluates six algorithmic and AI systems spanning finance, health, public services, employment, online commerce, and digital media. We measure the economically relevant dimensions of adoption and legitimacy, including beliefs, evaluative attitudes, and willingness to delegate decisions. Explanations yield measurable, though modest, increases in willingness to delegate. A mechanism-consistent decomposition shows that these effects arise primarily through improved attitudes toward the systems, while direct effects and belief shifts play a secondary role. Overall, explanations reliably move acceptance in the intended direction, but do not eliminate persistent concerns, especially those related to privacy. The results highlight both the promise and limits of information disclosure as a regulatory tool.