Syntes: Barcaui (2025) — ChatGPT as a cognitive crutch

Källa: Barcaui, A. (2025). ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention. Social Sciences & Humanities Open, 12, 102287. https://doi.org/10.1016/j.ssaho.2025.102287

Metadata

• ID: AI-010
• Tier: RCT
• N / design: RCT; 120 undergraduates randomised (AI-assisted ChatGPT GPT-4 web vs traditional non-AI study); 85 completed surprise retention test at 45 days
• Population: vuxen / HE (business undergraduates, Brazil)
• Subject: unrestricted ChatGPT as study aid vs traditional study; delayed closed-book retention

Finding

1. Traditional learners scored higher on the Day-45 surprise retention test than ChatGPT-assisted learners (M 6.85 vs 5.75 on a 10-point scale ≈ 68.5 % vs 57.5 %; t(83) = −3.19, p = .002, d = 0.68).
2. AI-assisted group spent less time on task (3.2 h vs 5.8 h); the AI disadvantage remained after controlling for study time (ANCOVA F(1,82) = 7.89, p = .006).
3. Prior AI familiarity did not significantly predict retention (r = 0.18, ns).

Mechanism

Cognitive offloading / removal of desirable difficulties: unrestricted ChatGPT reduces effortful encoding; retention measured without the tool is lower.

Limits

Single university, business programme, convenience sample; ~29 % attrition (balanced across arms); AI use in the control arm not directly monitored; presentation-preparation task may differ from exam-only study.

This claim dies if …

A replicated HE/K–12 RCT with delayed closed-book measurement finds typical unrestricted ChatGPT/answer-giving GenAI use does not lower retention vs control.

Swedish implication

Measure learning without the tool after practice with GenAI; unrestricted answer-giving use is a learning risk, not only a homework-quality boost.

Links

• DOI: https://doi.org/10.1016/j.ssaho.2025.102287
• Related: bastani-etal-2024-genai-harm-learning, fan-etal-2025-metacognitive-laziness, stromberg-lei-wu-2026-genai-learning-penalty