Wharton Study Finds Structured AI Improves Learning More Than On-Demand Help
New research from the Wharton School at the University of Pennsylvania suggests that artificial intelligence (AI) can significantly improve learning outcomes, but only when its assistance is structured rather than fully on demand.
The study, published on Tuesday (Feb. 24) by Knowledge at Wharton, examined how different forms of AI influence long-term skill development. In a three-month experiment involving more than 200 chess learners, researchers compared two approaches. One group received system-regulated AI guidance delivered at predetermined intervals; the other could request AI help whenever they chose.
Both groups had access to the same overall amount of AI assistance. The difference lay in when they could access it.
According to the study, learners who received structured, system-controlled guidance improved their performance by roughly 64% over the training period, while those with unrestricted, on-demand access improved by about 30%. Follow-up testing weeks later showed that the advantage of structured support persisted, indicating deeper retention rather than temporary gains.
The findings emphasize the importance of how AI is deployed.
Researchers attributed the stronger results in the structured group to what education psychologists describe as “productive struggle:” the process of working through challenging problems before receiving help. When learners tackled difficult positions independently and received guidance at calibrated moments, they were more likely to internalize strategies and strengthen decision-making skills.
Participants with immediate, self-directed access often resolved problems more quickly in the moment but engaged less deeply with underlying reasoning. Interviews revealed that many learners were aware that frequent assistance might limit their development, yet still opted for help when challenges arose. The study suggested that when support is frictionless and instantly available, self-regulation can falter.
The implications extend beyond chess training. As AI tutors and copilots become embedded in classrooms, professional training programs and enterprise learning platforms, the research offers guidance for developers and institutions. Systems that deliver calibrated prompts, staggered hints and bounded access to solutions may preserve the cognitive effort necessary for durable learning while still leveraging AI’s adaptive capabilities.
A related dimension of this broader conversation is captured in Anthropic’s AI Fluency Index, that measures how people are developing “fluency” with AI tools as they integrate them into everyday tasks.
The index tracks a taxonomy of behaviors—such as iteration and refinement, questioning the AI’s reasoning and specifying goals—that signal effective collaboration with AI systems. Early findings show that most productive interactions involve users building on previous exchanges with the model, rather than accepting first responses at face value.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
The post Wharton Study Finds Structured AI Improves Learning More Than On-Demand Help appeared first on PYMNTS.com.