AI Assistance Reduces Persistence and Hurts Independent Performance

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 (approximately 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.

• The paper demonstrates that short-term AI assistance improves immediate task success but impairs subsequent unaided performance and persistence.
• It uses three controlled experiments across mathematical reasoning and reading comprehension, revealing statistically significant declines in solve and persistence rates.
• The findings indicate that direct answer solicitation from AI leads to marked deskilling, highlighting the need for scaffolding over mere task completion.

AI Assistance and Its Detrimental Impact on Human Persistence and Independent Performance

Overview

The study "AI Assistance Reduces Persistence and Hurts Independent Performance" (2604.04721) provides a compelling, causal investigation into how short-term exposure to advanced AI assistance modulates human cognitive performance and motivation. By deploying a sequence of large-scale, carefully controlled behavioral experiments (total $N = 1,222$), the authors dissect the persistent and domain-general costs of AI-enabled cognitive offloading, with a particular focus on mathematical reasoning and reading comprehension. The results provide robust evidence that while AI assistance boosts immediate task performance, it significantly impairs subsequent unaided performance and reduces persistence, effects observable after only brief exposures and generalizable across cognitive domains.

Figure 1: AI impairs unassisted performance and persistence in fraction reasoning tasks; AI users experience a sharp drop in solve rate and an increase in skip rate once deprived of assistance.

Experimental Design and Methodology

Three sequential experiments structure the empirical investigation, each eliminating key methodological confounds and extending the findings:

• Experiment 1: A randomized controlled trial with fraction problems, comparing AI-assisted ($N = 185$) and non-assisted ($N = 122$) groups. AI assistance was available for 12 progressively difficult items, then silently withdrawn for a test phase.
• Experiment 2: A replication that resolves potential participant selection and interface confounds through a pretest and improved control sidebars; $N = 585$ post-exclusion.
• Experiment 3: A cross-domain extension to SAT-aligned reading comprehension problems, assessing generality.

In all cases, performance is assessed via solve (correct/attempt) rates and skip rates (volitional disengagement), with careful exclusion based on both attention and minimal skill thresholds.

Figure 2: Experiment 2 replicates the impairment in both performance and persistence, demonstrating statistical robustness and improved exclusion controls.

Main Results and Quantitative Findings

AI assistance caused a reliable and numerically significant deficit in unaided performance and persistence:

• Experiment 1: Post-assistance test problems yielded a mean solve rate of 0.57 (s.d. 0.41) for AI users vs 0.73 (s.d. 0.34) for controls ($t(305) = -3.64$, $P < 0.001$, Cohen's $d = -0.42$).
• Experiment 2: Even after skill-matching, AI users had a reduced test solve rate (0.71 vs 0.77, $t(583) = -2.33$, $P=0.020$, $d = -0.19$).
• Skip rates rose (indicative of decreased persistence): in Experiment 1, AI users skipped 20% vs 11% for controls ($N = 185$0, $N = 185$1).
• Localized effects: Costs are concentrated in users who solicited direct answers from the AI. Those who used AI only for hints exhibited performance comparable to or better than controls (see Figure 3).

  Figure 3: Persistence and performance decrements are isolated to participants who used AI for direct solutions, not merely for hints or clarification.

• Experiment 3 (Generalization): Reading comprehension tasks showed the same impairment (solve rate for AI group: 0.76, controls: 0.89; $N = 185$2, $N = 185$3, $N = 185$4), with the skip rate nearly an order of magnitude higher for AI users post-assistance.

  Figure 4: Reading comprehension persistence and performance both suffer similarly post-AI assistance.

Interpretation and Theoretical Implications

The findings align with and extend literature on cognitive offloading but go further by offering the first large-scale causal evidence that short-term AI assistance can rapidly erode both cognitive capability and motivational persistence in subsequent unaided work. Unlike classic external memory aids, LLM-based assistants can execute the full cognitive labor, thus altering the user's reference point for expected effort and denying exposure to productive struggle and metacognitive calibration. The effect is domain-general, manifesting in tasks spanning calculation and critical reading.

The observed pattern supports a model where:

• Immediate, unfiltered AI answers rapidly adjust subjective expectations, making unaided effort aversive.
• Metacognitive decay occurs, as feedback from actual problem-solving is replaced by solution consumption.
• Persistence is malleable and highly sensitive to assistance protocol, not just to underlying ability.

The stratification of effects by AI usage pattern further suggests that not all forms of assistance are equivalent: direct answer solicitation creates the most deleterious deskilling, unlike hint-based scaffolding, which appears relatively benign.

Broader Significance and Policy Implications

The deleterious effects observed emerge within a 10-minute exposure window, raising nontrivial concerns about cumulative effects of sustained or repeated AI use, especially in formative educational or skill-acquisition environments. These results challenge naive deployment of short-term optimization frameworks for human-AI collaboration, which ignore long-term user autonomy, self-efficacy, and the preservation of effort-regulation faculties.

Practical implications include:

• Design priorities for AI systems must shift toward scaffolding, not bare task completion. Human-AI interfaces should be optimized for long-term user competence.
• Policy: The findings necessitate revisiting deployment strategies in education and professional development, as superficial mitigation (e.g., time limits, user warnings) may not address the underlying incentive misalignment.
• Equity concerns: Deskilling effects will likely be concentrated among those without compensatory academic resources, potentially widening educational and cognitive disparities.

Future Directions

Longitudinal field deployment studies are warranted to quantify aggregate, population-level impacts and examine whether metacognitive decay and motivational atrophy persist, accumulate, or can be reversed by modifications in AI collaborative strategy (e.g., enforced Socratic tutoring, graduated autonomy). Further research should focus on differentiating effects by assistance modality and tuning the interface between immediate utility and long-term skill preservation.

Conclusion

The paper provides robust, multi-domain, and methodologically rigorous evidence that current paradigms for AI assistance yield significant and rapid erosion of unaided problem-solving capacity and volitional persistence. These findings substantiate a theory of gradual cognitive disempowerment induced by over-optimization for short-term helpfulness. They establish an empirical foundation for a paradigm shift toward human-empowering AI design, with long-term competence and autonomy as explicit objectives.

Figure 5: Schematic overview of experimental paradigm and core finding—AI assistance enhances immediate performance, but systematically impairs unaided persistence and capacity.