AI Assistance Reduces Persistence and Hurts Independent Performance (researchers at Carnegie Mellon, Oxford, MIT, UCLA)

[highplainsdem]

https://arxiv.org/html/2604.04721v1

Abstract
People often optimize for long-term goals in collaboration: A mentor or companion doesn’t just answer questions, but also scaffolds learning, tracks progress, and prioritizes the other person’s growth over immediate results. In contrast, current AI systems are fundamentally short-sighted collaborators – optimized for providing instant and complete responses, without ever saying no (unless for safety reasons). What are the consequences of this dynamic? Here, through a series of randomized controlled trials on human-AI interactions (N = 1,222), we provide causal evidence for two key consequences of AI assistance: reduced persistence and impairment of unassisted performance. Across a variety of tasks, including mathematical reasoning and reading comprehension, we find that although AI assistance improves performance in the short-term, people perform significantly worse without AI and are more likely to give up. Notably, these effects emerge after only brief interactions with AI (∼10 minutes). These findings are particularly concerning because persistence is foundational to skill acquisition and is one of the strongest predictors of long-term learning. We posit that persistence is reduced because AI conditions people to expect immediate answers, thereby denying them the experience of working through challenges on their own. These results suggest the need for AI model development to prioritize scaffolding long-term competence alongside immediate task completion.

-snip-

6 Conclusion
Human cognition has always been shaped by external tools, from calculators to internet to GPS navigation. Current AI systems, however, represent a new kind of cognitive scaffold: one that solves anything, rarely refuses to help, and delivers answers instantly. Here, we show that just 10–15 minutes of AI interaction can result in significant impairments in independent performance and persistence – capacities that are foundational to life-long learning. If brief exposure produces measurable erosion, the cumulative effects of daily AI use over months or years may be profound and difficult to reverse.

Two mechanisms may explain the observed decline in persistence. First, when AI routinely completes tasks in seconds, the reference point for how long a task should take can shift – and as a consequence, unaided work starts to feel counterfactually more effortful, a process structurally analogous to hedonic adaptation (Brickman, 1971; Brickman et al., 1978; Frederick & Loewenstein, 1999). Crucially, this mechanism is self-reinforcing: each act of offloading shifts the reference point, increases the subjective cost of unaided effort, and makes future offloading more attractive. Second, AI removes the productive struggle through which people develop not only accurate knowledge but accurate self-knowledge. Without opportunities to work independently, people never learn what they are capable of, undermining the metacognitive calibration that sustains persistence (Yeung & Summerfield, 2012; Fleming & Daw, 2017; Dubey et al., 2021; Elizondo et al., 2024). Together, these mechanisms predict broader metacognitive decay beyond persistence alone – a direction future work should examine in naturalistic, longitudinal settings.

Our results carry important policy implications. The tasks investigated here, such as fraction arithmetic and reading comprehension, may seem delegable to tools like calculators, but conceptual mastery of these skills is a developmental prerequisite. Without these skills, higher-order competencies like algebra or critical reasoning remain inaccessible. If sustained AI use erodes the motivation and persistence that drive long-term learning, these effects will accumulate over years, and by the time they are visible, they will be difficult to reverse. This is analogous to the “boiling frog” effect, where each incremental act feels costless, until the cumulative effect becomes overwhelming to address (Moore et al., 2019; Kasirzadeh, 2025; Liu et al., 2025). These risks are also not equally distributed: students with fewer academic resources may be most vulnerable. While user-facing interventions (e.g., Socratic AI, reduced use time, etc.), might help at the margins, we believe that such solutions will only serve as “band-aids” and will not resolve the deeper issue, since AI offers a temptation to offload at scale. Mitigating these risks requires rethinking how AI systems collaborate with people long-term and broadening objectives beyond short-term user satisfaction (Zhi-Xuan et al., 2025; Kirk et al., 2025b) toward an ethos of empowerment and care (Kleiman-Weiner, 2024; Christian, 2025). We hope that our work inspires the field to think about optimizing not just what people can do with AI, but what they can do without it.

-snip-