Do forcing functions reduce AI overreliance without harming team performance?

An experiment with N=199 participants on Amazon Mechanical Turk, using a simulated AI assisting a nutrition-substitution decision task, finds that cognitive forcing functions — interventions "applied at the decision-making time to disrupt heuristic reasoning and thus cause the person to engage in analytical thinking" — significantly reduce overreliance on AI compared with two simple explainable-AI (XAI) designs. But the reduction in overreliance does not translate into higher overall human+AI performance: "even with the cognitive forcing functions, the human+AI teams in our study continued to perform worse than the AI model alone," and there was "no significant difference in performance between simple explainable AI approaches and cognitive forcing functions" — the hypothesis that forcing functions would improve team performance overall (H1b) was not supported, only the narrower hypothesis that they reduce overreliance specifically when the AI is wrong (H1a). The trade-off cuts against usability too: "participants assigned the least favorable subjective ratings to the designs that reduced the overreliance the most," and when the simulated AI was wrong, "people performed best in conditions that they preferred and trusted the least."

The paper grounds the overreliance problem in dual-process theory: people mostly rely on fast System 1 heuristics rather than effortful System 2 analysis, and "evaluating every explanation requires substantial cognitive effort, which humans are averse to" — so explanations function as "a general signal of competence" that can increase trust and overreliance rather than being read individually for their content. Two of the three forcing functions tested were "on demand" (the AI suggestion is hidden until the user clicks to reveal it) and "update" (the person commits to an initial answer before seeing the AI's suggestion, then may revise it) — both designed to insert a moment of required engagement before the AI's answer is accepted, rather than changing what the AI shows.

This sits alongside Does AI assistance always help reasoning or does it carry hidden costs?: both treat intervention friction as the lever, but where that note frames friction as a cost to be minimized, this study shows the friction itself — forcing the pause — is the active ingredient that reduces overreliance, at the price of lower acceptability. It also qualifies Can AI guidance reduce anchoring bias better than AI decisions?: where that note's LTG approach redesigns what the AI shows to avoid anchoring, this study shows that even holding the AI's output constant, simply gating when and how the suggestion is revealed changes how much people anchor on it.

The task is a low-stakes, simulated proxy domain (swapping high-carbohydrate ingredients in a meal image) with a simulated AI set to 75% accuracy, run on Mechanical Turk — not a clinical, financial, or other high-stakes real-world decision; the paper itself notes that "accessing experts such as judges, or clinicians for experiments is notoriously challenging and costly," which is why it substituted nutrition judgments. The excerpt does not establish that cognitive forcing functions ever close the gap to AI-alone performance, only that they narrow the overreliance component of it; whether friction-based designs are viable outside a lab setting, where users could simply abandon a cumbersome interface rather than tolerate it, remains open.