Source-linked AI summary
Distinguishing performance gains from learning when using generative AI
Lixiang Yan, Samuel Greiff, Jason M. Lodge, Dragan Gašević
TL;DR
The paper addresses whether generative AI’s educational performance gains represent durable learning, given risks to cognitive and metacognitive processing. It synthesizes evidence on cognitive load, metacognition, autonomy and self-efficacy, and proposes assessments and practices that prioritize retention, transfer and independent learning. Overall, AI should augment rather than replace human cognitive effort.
Problem
Generative AI can improve immediate task performance and reduce cognitive load, but may undermine the cognitive and metacognitive processes required for durable learning.
Source: Main · p. 2, Cognitive load, metacognition and… · p. 3Method
The paper examines cognitive and metacognitive mechanisms and proposes process-based, longitudinal research alongside educational practices that preserve active learner engagement.
Source: Targeting learning in research and… · pp. 3–4Results
Generative AI is associated with lower cognitive load and higher perceived confidence, but also weaker reasoning, metacognitive engagement, autonomy and independent learning.
Source: Cognitive load, metacognition and… · pp. 2–3Takeaways & Limitations
Educational AI should support, not replace, independent cognitive effort, with retention, transfer and unaided problem-solving used to distinguish learning from task success.
Source: Targeting learning in research and… · p. 4, Outlook · p. 4
Abstract
from arXiv · showhide
Generative artificial intelligence (AI) is increasingly being integrated into education, where it can boost learners' performance. However, these uses do not promote the deep cognitive and metacognitive processing that are required for high-quality learning.
Main
Generative AI can improve learners’ immediate task performance, but performance gains do not necessarily indicate durable learning. Emerging research often conflates these outcomes, leaving the distinction underexamined.
- Generative AI can boost performance and reduce cognitive load, but may undermine cognitive and metacognitive processes required for durable learning.Source: Main · p. 2
- Learning entails enduring, independently retained and transferable knowledge or skills, whereas performance is observable task behaviour shaped by external supports.Source: Main · p. 2
- A meta-analysis of 69 studies reported improved academic performance with Hedges’ g = 0.7, potentially reflecting immediate task success rather than learning.Source: Main · p. 2
Cognitive load, metacognition and self-efficacy
Generative AI can lower cognitive load and increase confidence, yet offloading cognitive and metacognitive work may weaken reasoning, autonomy, motivation and independent learning.
- Generative AI can reduce cognitive load by offloading complex information-processing tasks, but excessive offloading may reduce active engagement needed for retention and transfer.Source: Cognitive load, metacognition and… · pp. 2–3
- Heavy ChatGPT reliance for information gathering was associated with weaker argumentation and reasoning than traditional, cognitively demanding research methods.Source: Cognitive load, metacognition and… · p. 3
- Frequent generative AI use can foster metacognitive laziness by shifting evaluation, reflection and self-regulation tasks from learners to the tool.Source: Cognitive load, metacognition and… · p. 3
- Frequent chatbot interaction was associated with lower learning autonomy, while offloading planning, monitoring and evaluation may reduce intrinsic motivation.Source: Cognitive load, metacognition and… · p. 3
- Generative AI can increase perceived confidence and efficiency while also increasing technological dependence, potentially weakening independent learning and resilience.Source: Cognitive load, metacognition and… · p. 3
Targeting learning in research and practice
The paper proposes research and educational practices that distinguish learning from performance and examine the cognitive mechanisms and long-term effects of generative AI use.
- Research should examine how generative AI affects encoding, consolidation and retrieval, including the relationship between cognitive load and deep encoding.Source: Targeting learning in research and… · pp. 3–4
- Understanding underlying cognitive mechanisms can clarify whether AI assistance promotes deep engagement, knowledge transfer and durable skill development rather than task completion.Source: Targeting learning in research and… · p. 4
- Researchers should use retention, transfer and delayed-recall assessments to distinguish durable learning from immediate task performance.Source: Targeting learning in research and… · p. 4
- Educators should encourage critical evaluation of AI outputs and balance efficiency with learner autonomy and independent cognitive effort.Source: Targeting learning in research and… · p. 4
- AI feedback can be combined with unaided problem-solving to support learner autonomy alongside task success.Source: Targeting learning in research and… · p. 4
Outlook
Generative AI may improve performance and reduce cognitive load while disrupting processes needed for durable learning. The paper argues that AI should augment, not replace, human learning.
- Generative AI can enhance performance and reduce cognitive load but might disrupt encoding, retention and independent problem-solving.Source: Outlook · p. 4
- Process-based assessments, cognitive-function analysis and long-term evaluation can support a more nuanced understanding of human–AI interactions.Source: Outlook · p. 4
- Generative AI should augment human learning so learners develop knowledge and skills beyond AI-assisted tasks.Source: Outlook · p. 4