Episode Overview
Part 2 of this 3-part series examines whether general “thinking skills” exist independently of subject matter, and measures what happens to the human brain when the friction of thinking is offloaded to AI.
Tracing 120 years of cognitive science—from Thorndike’s refutation of Latin grammar training to chess grandmaster memory chunking—we contrast the science of expertise with recent 2024–2026 studies on AI offloading, retention drops, and persistence collapse.
Core Research Highlights
1. The Myth of Far Transfer
Employers routinely rank “critical thinking” as a top skill, but second-order meta-analyses (Sala & Gobet, 2019) show that cognitive training in chess, music, or working-memory games produces an effect size of exactly zero for general far transfer. General thinking skills are almost entirely domain knowledge in disguise.
2. Expertise as Pattern Recognition
Adriaan de Groot’s landmark chess studies revealed that grandmasters reconstruct real game boards with 93% accuracy after 5 seconds, but their memory advantage drops to zero when pieces are placed randomly. Expertise relies on stored functional “chunks” built through effortful, hands-on practice.
3. The AI Offloading Paradox & Cognitive Debt
- Bastani et al. (2023/2025 PNAS): High school students using an AI math tutor saw a 127% boost in practice performance, but scored 17% worse than the control group on unassisted exams.
- Barcaui (2025): Students using ChatGPT finished 2 hours faster, but suffered an 11 percentage point drop on a surprise retention test 45 days later.
- Liu et al. (2026 CMU Study): Just 10 minutes of using AI direct answers doubled participants’ skip rate on subsequent unassisted tasks (20% vs. 11%), showing a rapid collapse in psychological persistence.
Key Takeaways
- Evaluating Requires Producing: Evaluating an AI’s answer requires the exact same domain knowledge built by producing answers yourself.
- Cognitive Debt Accumulation: Skipping the cognitive friction of manual work deprives the brain of the mental templates needed for oversight.
- Persistence Degradation: Relying on frictionless AI answers rapidly erodes persistence on difficult unassisted problems.
References & Cited Research
- Thorndike, E. L., & Woodworth, R. S. (1901). The Influence of Improvement in One Mental Function upon the Efficiency of Other Functions.
- De Groot, A. (1946/1965). Thought and Choice in Chess.
- Sala, G., & Gobet, F. (2019). Near and Far Transfer in Cognitive Training: A Second-Order Meta-Analysis.
- Bastani, H., et al. (2023/2025). Generative AI Without Guardrails Can Harm Learning. PNAS.
- Barcaui, A. (2025). ChatGPT and Long-Term Memory Retention. Social Sciences & Humanities Open.
- Liu, G., Christian, B., et al. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. CMU / arXiv.