Tomu Tominaga, Naomi Yamashita, Takeshi Kurashima
5 min
Algorithmic recourse provides users with counterfactual "action plans" to overturn unfavorable AI decisions (e.g., a loan rejection). While researchers often suggest that providing a diverse set of options helps users find more feasible paths and increases motivation, it also requires more complex counterfactual reasoning. This study investigates the trade-off between the psychological benefits (e.g., willingness to act) and psychological costs (e.g., cognitive load, negative emotions) of diversifying these recourse sets.
The authors conducted a controlled, between-subjects experiment with 750 participants using a hypothetical car loan application scenario. They manipulated two variables:
Participants were asked to evaluate the recourse sets based on subjective reasonability, actionability, willingness to act, and decision acceptance (benefits), as well as cognitive load and emotional burden (costs).
The study reveals a clear "sweet spot" for diversifying algorithmic recourse:
This research demonstrates that more is not always better when it comes to AI explanations. For developers and designers, the findings suggest that diversification is highly effective for small sets (e.g., three options), as it boosts user motivation without overwhelming them. However, when providing larger sets, designers must implement safeguards or better filtering methods to prevent the "cognitive burden" of counterfactual reasoning from undermining the user's experience. The study highlights the necessity of incorporating human psychological constraints into the design of explainable AI systems.
Algorithmic recourse provides counterfactual action plans that help people overturn unfavorable AI decisions. While diverse recourse sets may improve transparency and motivation, they may also impose cognitive load and negative emotions by increasing counterfactual reasoning demands. To examine this trade-off, we conducted a between-subjects controlled experiment (N=750) that manipulated recourse-set diversity and size, and evaluated these effects on psychological benefits and costs. Results show that diversification enhances psychological benefits (e.g., willingness to act) for small sets without incurring additional psychological costs, whereas for large sets, it makes cognitive load more salient. These findings suggest that naively diversifying recourse can burden decision subjects, underscoring the need for new diversification methods that incorporate human cognition and psychology to mitigate such costs.
Alex: Can you give me a concrete example?
Sam: Sure. Imagine the AI suggests that one way to get a loan is to take on a side job to increase your income. But then, to give you a "diverse" second option, it suggests you should *quit* your current side job to reduce your debt load. The system is telling you to move the same needle in opposite directions across different plans.
Alex: So if you're the applicant, you're looking at those two plans thinking—does the AI want me to work more or less? The whole logic starts to feel like a coin toss.
Sam: Precisely. And once the advice feels contradictory, people stop trusting it. The sense of being in control—of having a clear path forward—disappears, and people become less likely to act at all.
Alex: So where does that leave the "sweet spot"? Is three options the right number?
Sam: That seems to be the case in this study. At three options, diversity was a genuine benefit—it boosted how motivated people actually felt to follow the advice. But once they hit seven options, that motivation plateaued. Meanwhile, the mental effort required to just *understand* the advice kept climbing. The group receiving seven diverse options reported more mental strain than the group receiving seven very similar ones.
Alex: So the takeaway for people building these systems is that "more" isn't a strategy. You have to balance variety with the mental cost of processing it.
Sam: Right. The researchers also point out that the AI needs rules for what it can suggest in the first place. If a person's age is a factor in the decision, the AI shouldn't suggest "becoming younger" as a path to approval. The advice has to stay within what's actually feasible for that specific individual.
Alex: Did the study find any other hidden traps in those larger sets?
Sam: They did look at whether having a "common change"—where every single option tells you to do the exact same thing—made people feel frustrated, like they weren't really getting a choice. Interestingly, that didn't bother people nearly as much as the contradictions did. The real tipping point for mental exhaustion was the inconsistency between paths, not the repetition within them.
Alex: That makes sense. I'd rather be told the same helpful thing three times than be told two opposite things once.
Sam: Right. So the study suggests that future AI systems might need to actively track how many contradictions they're generating. If the math starts producing too many conflicting directions, the system should probably limit the options shown—rather than presenting every path and leaving the person to sort out the mess.
Alex: It's a meaningful shift in thinking—from "more data is better" to "less is more, if it's consistent." At the end of every algorithm, there's a person trying to make a real decision about their life.
Sam: And that's what this research keeps coming back to. Giving someone more control only works if the advice they receive actually makes sense. When the logic feels broken, people are less likely to act at all. That's worth keeping in mind as these systems become more common.
Alex: Thanks for listening to ResearchPod.