AI's classroom shortcut can produce the work while students lose the struggle that builds thought

The essay is a warning, not a neutral review

The Guardian piece argues that reading, writing, struggle, and delay are not inefficiencies to remove from education. They are the mechanisms through which students build stamina, judgment, memory, and an independent voice. It warns that a generated essay can look like mastery to a student, teacher, or institution even when the learner never formed the understanding behind it.

That argument should be labeled correctly. It is an opinion essay grounded partly in classroom observation and selected research. It does not demonstrate that every use of generative AI causes cognitive harm, nor does it compare structured tutoring systems with unrestricted answer generation. Its value is to force a distinction that school policy often evades: completing an assignment and building a capability are different outcomes.

Early evidence supports caution, not panic

A 2026 preprint reports randomized experiments across 1,222 participants completing mathematics and reading-comprehension tasks. AI assistance improved short-term performance, but participants performed worse and gave up more frequently after losing access. The effects appeared after roughly 10 to 15 minutes of assistance. The study directly tests persistence and unassisted performance, but it does not establish the effects of years of classroom use or prove the same magnitude for children.

A separate MIT Media Lab preprint studied 54 adults across AI-assisted, search-assisted, and unassisted essay-writing conditions, with 18 returning for a fourth session. The LLM group showed weaker EEG connectivity and lower recall and ownership of their essays. Its small sample, preprint status, and debated EEG interpretation limit broad claims. Together, the studies justify measurement and safeguards, not declarations that a generation has already been cognitively damaged.

The learning process is the product

A school can raise apparent output while reducing practice. If a model drafts the argument, selects evidence, supplies the phrasing, and repairs the logic, the student may submit a better artifact while rehearsing fewer of the skills the assessment was designed to build. The educational risk is therefore not only cheating. It is curriculum that quietly mistakes assisted fluency for independent competence.

That risk is unequal. Well-resourced schools can preserve human teaching, oral discussion, books, laboratories, and supervised practice while using AI selectively. Under-resourced systems may be pressured to replace expensive human support with cheaper automated assistance. A tool sold as personalized learning could deepen the divide between students taught to think and students trained to accept an answer.

What schools should measure now

The useful policy question is not whether AI belongs everywhere or nowhere. It is which cognitive objective an assignment serves and whether the student can demonstrate that objective without hidden assistance. Schools need unassisted transfer tasks, oral explanation, source checks, version history, and delayed recall alongside any productivity gains reported by an AI tool.

A good tutor withholds answers when productive struggle matters. Classroom AI should be judged by the same standard. If it maximizes immediate completion while weakening persistence or independent performance, it has optimized the wrong educational outcome.

• Measure delayed recall and unassisted transfer, not only assisted completion.
• Require students to explain decisions and correct model errors in their own words.
• Differentiate answer generation from scaffolded tutoring in policy and procurement.
• Protect human teaching time before using AI as a substitute for staff.