Distinguishing performance gains from learning when using generative AI

Lixiang Yan, Samuel Greiff, Jason M. Lodge, Dragan Gašević (2026) — Nature Reviews Psychology, 4(7), 435-436.

📄 Full text (arXiv) | Published

Core Argument

This Nature Reviews Psychology piece draws a critical distinction that has been under-theorized in AIED research:

• Performance gains — immediate improvements in task completion, efficiency, or output quality when using AI tools.

• Learning — durable understanding that requires deep cognitive processing (elaboration, critical analysis) and metacognitive processing (planning, monitoring, reflection).

The authors argue that generative AI easily boosts performance but often bypasses the mental processes essential for genuine knowledge acquisition. This challenges the common assumption in ai-tutor-effectiveness-review research that improved task performance equals improved learning.

Implications for AIED Design

This distinction has profound implications for scaffolding design. Systems that optimize for immediate performance may undermine learning. They must be designed to promote cognitive engagement — for example, through socratic-method dialogue, constrained feedback, or requiring student articulation before revealing AI output. The piece connects to over-reliance research showing that AI assistance can reduce independent problem-solving and to the pedagogy-ai-mistakes paradigm that uses AI errors as learning opportunities.

The performance-vs-learning gap is now causally demonstrated in a field RCT: generative-ai-guardrails-harm-learning shows that an unguarded GPT-4 tutor raised practice performance +48% but reduced later unassisted exam scores by 17%, while a guardrailed "hint-not-answer" tutor eliminated the harm.

Theoretical Framework

Generative AI can function as either a cognitive tool (amplifying thinking) or a cognitive crutch (replacing thinking). This maps onto the self-regulated-learning cycle — performance gains without metacognitive engagement short-circuit the planning-monitoring-evaluating loop.

Related Pages

• ai-assistance-reduces-persistence: Causal evidence (N=1,222) that brief AI assistance reduces persistence and impairs unassisted performance — rapid emergence of over-reliance effects

• cognitive-offloading-speedup-illusion — Speedup illusion parallels performance-vs-learning distinction in GenAI use

• cognitive-shift-ai-education — 471 students surveyed 2020–2026 show shift from AI preference to human intellige

• students-llm-usage-critical-thinking — LLM usage patterns and actual learning outcomes

• taklif-ai-interest-based-personalized-assignments — Engagement gains need learning outcome validation

• eduframetrap-llm-sycophancy-educational-safety — Agreeableness undermines corrective friction needed for learning

• ai-learning-transfer — Evidence that AI-boosted performance may not transfer

• metacognition — Cognitive processes that generative AI can bypass

• self-regulated-learning — SRL cycle disrupted by performance-only AI use

• over-reliance — AI as cognitive crutch rather than tool

• scaffolding — Designing AI scaffolds that promote rather than replace thinking

• ai-learning-companions-framework — Prioritizing learning over performance

• learning-by-chatting-genai-impact — ChatGPT improved performance on search tasks but reduced actual learning outcomes

Citation

APA: Yan, L., Greiff, S., Lodge, J. M., & Gašević, D. (2026). Distinguishing performance gains from learning when using generative AI. Nature Reviews Psychology, 4(7), 435-436. arXiv:2605.13731.