Chapter 5 — AI and the Mind: Cognition, Critical Thinking, and Dependence
Last updated: July 5, 2026
The deepest question in the book is not whether AI raises this year's grades but what sustained use does to thinking itself — to memory, reasoning, persistence and the disposition to do hard cognitive work. Education's product is not test scores; it is minds, and a technology could lift grades while quietly eroding the capacities the grades are meant to measure. Public anxiety concentrates here: two-thirds of English secondary teachers believe pupils' critical thinking has already declined because of AI. The honest position is watchful uncertainty, not settled alarm — and the chapter's answer is a single precise sentence: the danger is not using AI, it is letting AI do the part of the task that was the point of the task.
The science of cognitive offloading predates ChatGPT by decades — Socrates warned in the Phaedrus that writing would breed forgetfulness — and is reassuring in one respect. Risko and Gilbert's review established that offloading is often rational and performance-enhancing; the trouble is that we offload more than is optimal when the tool is convenient, and what we offload we tend not to encode. Nobody mourns the mental arithmetic lost to calculators, because curricula decided, fight by fight, which fluencies still mattered and taught them deliberately. But the costs are specific rather than general: you lose what you stop practising, roughly in proportion to how completely the tool absorbs the practice. The AI question is whether it absorbs practice on something education cannot afford to stop building — the construction of understanding itself.
The best single piece of evidence is the Bastani, Bastani and Sungu randomised trial in Turkey. Students given unrestricted ChatGPT during maths practice performed 48 percent better on the practice problems and 17 percent worse on the subsequent unassisted exam; the usage logs showed them using the bot to get answers rather than help, outsourcing the effortful retrieval that produces durable learning. The guarded version that offered hints but withheld answers eliminated the harm — the mechanism, not the tool, was the problem. Boundary conditions matter, and they cut both ways: a separate randomised study of 334 university adults preparing for an incentivised exam (Chapter 4's Fischer result) found the reverse, with unrestricted access beating gated access by 0.21 standard deviations with logs showing self-regulated use. The crutch bites hardest where the chapter expects it to: younger learners acquiring new skills, not motivated adults consolidating material.
A further study gave the pattern a name. Fan, Tang and colleagues, in the British Journal of Educational Technology, randomised learners to write with ChatGPT, with a human expert, with a writing-analytics tool, or unaided, and traced their processes rather than only their outputs. The ChatGPT group improved immediate scores but offloaded the self-regulatory work — planning, monitoring, reflecting — onto the tool, showing no advantage in knowledge gain or transfer. They called it metacognitive laziness — offloading theory playing out in a field experiment.
The viral brain-rot studies prove far less than their headlines. MIT's cognitive-debt paper (Kosmyna and colleagues) had participants write essays with an LLM, a search engine, or unaided while wearing EEG headsets; the LLM group showed the weakest connectivity, reported the least ownership, and often could not quote essays written minutes earlier. But it rests on 54 participants, only 18 in the crucial crossover session, and remains an unreviewed preprint — and EEG connectivity during a task is not a measure of learning. The authors themselves objected to the dumber headlines. Lee and colleagues' Microsoft/CMU survey of 319 knowledge workers found trust in the AI substituting for scrutiny of it, but it is self-report and correlational. The much-cited Gerlich correlation, across 666 participants, is cross-sectional, received a published correction, and cannot support the causal story pinned to it.
The pen-versus-keyboard saga is the cautionary parallel: Mueller and Oppenheimer's famous longhand advantage failed two direct replications — a reminder that single famous studies are weak ground for policy. And the other ledger exists — a CHI 2025 trial found an AI tool that coached students on a predicted grade lifted learning gains 8.9 percent over control, helping the overconfident most. AI that does the thinking harms; AI that provokes thinking can help.
The practical translation is therefore exact, and it is the most important instruction in the book: identify the cognitive move each assignment exists to build — retrieval in a vocabulary drill, argument construction in an essay, planning in a project — and protect that specific move from offloading. That is precisely what the guardrails and task designs of Chapter 11 are engineered to do.
Sources. [Primary empirical] Bastani, H., Bastani, O., Sungu, A., et al. "Generative AI Without Guardrails Can Harm Learning." PNAS (2025). doi:10.1073/pnas.2422633122. https://www.pnas.org/doi/10.1073/pnas.2422633122 · [Primary empirical] Kosmyna, N., et al. "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." Preprint (2025; v2 December 2025). arXiv:2506.08872. https://arxiv.org/abs/2506.08872 (project page: https://www.media.mit.edu/publications/your-brain-on-chatgpt/; authors' FAQ: https://www.brainonllm.com/faq; commentary: https://arxiv.org/pdf/2601.00856; measured coverage: https://www.techlearning.com/news/your-brain-on-chatgpt-everything-educators-need-to-know-about-mits-ai-study) · [Primary empirical] Fan, Y., Tang, L., et al. "Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance." British Journal of Educational Technology 56 (2025). doi:10.1111/bjet.13544. https://bera-journals.onlinelibrary.wiley.com/doi/abs/10.1111/bjet.13544 (preprint: https://arxiv.org/abs/2412.09315)