Does explaining AI decisions actually reduce user overreliance on it?

Giuseppe Romeo and Daniela Conti's systematic review, following PRISMA 2020 guidelines across 35 peer-reviewed studies drawn from SCOPUS, ScienceDirect, PubMed, and Google Scholar and published between January 2015 and April 2025, finds that explainable AI (XAI) and transparency mechanisms — tools meant to show users why an AI system reached a recommendation — frequently fail to curb automation bias (AB), "the tendency to over-rely on automated recommendations," and can instead deepen it. The review states that "although explanations may increase perceived system acceptability, they are often insufficient to improve decision accuracy or mitigate AB," and that "overly technical, cognitively demanding, or even simplistic explanations may inadvertently reinforce misplaced trust, especially among less experienced professionals with low AI literacy." Current XAI methods "have often shown that explanations themselves can be used as heuristics" and "reinforce trust in AI systems, exacerbating AB when the advice provided is incorrect."

The review's own synthesis points to user engagement, not explanation design alone, as the lever that works: "increased verification effort has been shown to reduce complacency toward AI mis-recommendations," so the authors propose "explanation design strategies that actively promote critical engagement and independent verification" rather than strategies aimed only at making systems more transparent or acceptable. The review frames this as a design trade-off: "an effective XAI paradigm must balance acceptability (building trust) with skepticism (encouraging critical evaluation)." It also identifies AI literacy and professional domain expertise as moderators that reduce AB, while high workload, time pressure, and task complexity reinforce it, and notes that the sequencing of advice matters — AI-first protocols can raise compliance, while human-first protocols can produce automation complacency when the AI's suggestion aligns with a person's initial judgment.

This generalizes a mechanism already visible in single studies in the library. Do forcing functions reduce AI overreliance without harming team performance? shows one specific engagement-forcing design — hiding or delaying the AI's suggestion — cutting overreliance at the cost of being the least-liked option, which matches this review's broader claim that effective debiasing tends to cost acceptability. Does access to web search prevent overreliance on chatbots? offers a parallel finding at the tool level: merely giving users more resources to check an answer does not help if verification itself is driven by trust rather than correctness — consistent with this review's claim that explanation availability alone is not the active ingredient. Do users worldwide trust confident AI outputs even when wrong? matches this review's account of perceived reliability (confidence, apparent objectivity) as a driver of AB independent of whether the underlying recommendation is right.

The review synthesizes findings from 35 cited studies rather than running new experiments itself, so its claims inherit whatever limitations those underlying studies carry (most are described as lab or simulated-task settings) and the review does not quantify how much explanation design typically increases versus decreases accuracy across contexts. It also does not resolve a tension it itself raises: a subset of the reviewed studies argues that what looks like automation bias in experts is actually confirmation bias — experts accepting AI advice that matches their prior judgment and rejecting advice that does not — which the review addresses only by noting those studies fail to account for cases where users defer to AI despite expressed reluctance. The implication the review supports is that interaction design aimed at AB should target verification behavior directly rather than assuming more or clearer explanation will do the job.