How do anchoring and Dunning-Kruger deepen AI overreliance?

Giuseppe Romeo and Daniela Conti's systematic review of 35 studies on automation bias (AB) — overreliance on automated recommendations — reports that AB "frequently interacts with other cognitive biases, particularly the anchoring effect, where initial AI recommendations disproportionately shape subsequent decisions." The review states this "leads decision-makers to seek evidence that confirms the AI initial suggestion," citing Gaube et al. (2021) and Nourani et al. (2022), "especially when incorrect AI guidance is provided before they form their own judgment," citing Agudo et al. (2024). Separately, the review links AB "to the Dunning–Kruger effect, where individuals with limited AI knowledge overestimate their understanding of it, which makes them more prone to overreliance on automation," citing Horowitz and Kahn (2024).

In the review's account, these are distinct mechanisms that both push in the direction of more reliance, not less. Anchoring operates through the order and framing of advice: when an AI's suggestion arrives before a person has formed an independent judgment, that suggestion becomes the reference point subsequent reasoning confirms rather than tests. The review notes the same anchoring logic can flip into "automation complacency" when the AI's recommendation, offered as a second opinion, happens to align with what the person already believed — the bias looks different depending on sequence, but the result is the same reduced scrutiny. Dunning-Kruger operates through self-assessment: people who know the least about how an AI system works are the ones least able to judge when to doubt it, which the review distinguishes from AI literacy as a protective factor — literacy reduces AB directly, while the Dunning-Kruger pattern describes why low-literacy users do not recognize their own exposure to it.

This sits alongside Does explaining AI decisions actually reduce user overreliance on it?, the sibling note from this review: that note covers the review's claim that explanations and transparency often fail to fix AB, while this note covers the review's account of why AB persists even when people are not simply trusting a system blindly — anchoring and self-overestimation are active cognitive mechanisms, not mere absence of information. It also connects to Do users worldwide trust confident AI outputs even when wrong?: that study's finding that confidence signals dominate accuracy tracking is one plausible channel through which anchoring operates, since a confidently stated first answer is exactly the kind of initial recommendation the review describes becoming the anchor. And it bears on Does unmodified chatbot behavior block rule discovery through sampling bias?, where confirmatory sampling toward an initial hypothesis parallels the confirmatory evidence-seeking the review attributes to anchoring.

The review itself flags a limitation worth carrying forward: a subset of the studies it covers (Alon-Barkat and Busuioc 2023; Selten et al. 2023; Bashkirova and Krpan 2024) argues that what looks like AB in experts may instead be confirmation bias — experts accepting AI advice that matches their prior judgment and resisting advice that does not — which would locate the mechanism in the expert's own prior belief rather than in anchoring to the AI's suggestion per se. The review responds only that those three studies "fail to account for cases where participants rely on incorrect AI recommendations despite their expressed reluctance." Both the anchoring and Dunning-Kruger claims here are drawn from citations within the review rather than from experiments the review ran itself, so the strength of each claim rests on the cited primary studies, not on independent replication by Romeo and Conti.