Three years ago, ChatGPT launched and detonated inside the education system like a slow-motion bomb. Teachers began receiving essays that were grammatically perfect, structurally coherent, factually adequate — and hollow. School districts scrambled to update plagiarism policies. Turnitin deployed AI detection tools. Entire schools banned AI outright. Others declared they’d embrace it.
In 2026, most of the emergency has passed and a fragile consensus has emerged: AI is here, students will use it, and somehow we need to teach alongside it. But the harder question — the one that doesn’t get answered by an acceptable use policy — is what happens to the mind when a powerful tool takes over the work that would have built it.
What Writing Actually Does (That AI Bypasses)
The research on writing and cognitive development is unambiguous: the act of writing is not just the output of thought, it is a driver of thought. When you are forced to articulate something in writing — to decide on a thesis, construct an argument, find the right word for a complex idea — you are doing something cognitively irreplaceable. The difficulty is the point.
Writing forces you to:
- Identify the gaps in your understanding (you can’t write clearly about something you only half-understand)
- Organize ideas into logical sequences, which requires actually thinking through the logic
- Synthesize information from multiple sources into a coherent perspective
- Discover, through the act of writing, what you actually think about something
When a student hands an essay prompt to ChatGPT, reads the output, makes minor edits, and submits it — none of that happens. The model has done the thinking. The student has done the editing. These are different cognitive activities, and only one of them builds the capacity for sustained original thought.
This isn’t a moral argument about academic honesty. It’s a neuroscience argument about what activities build cognitive capability and which ones don’t.
The Retrieval Practice Problem
Cognitive science research on learning has established that retrieval — the act of pulling information out of memory — is one of the most powerful drivers of long-term retention. The “testing effect” is one of the most replicated findings in educational psychology: being tested on material, even before you’ve learned it well, produces better long-term retention than equivalent time spent re-reading or passively reviewing.
AI tools, used carelessly, eliminate retrieval. Why memorize the dates of key historical events if you can ask your phone? Why remember how to structure a persuasive argument if the AI will generate one for you? Why learn to solve problems step-by-step if the answer appears instantly?
The argument “we don’t need to memorize things because we can always look them up” predates AI — it’s the same argument that was made about calculators, search engines, and Wikipedia. But the research consistently shows that having information in your head — actually knowing things — matters for the higher-order thinking we say we want students to be capable of.
You can’t reason from premises you have to look up every time you need them. Fluent expertise in any domain requires a large base of internalized knowledge that you don’t have to retrieve from an external source. AI tools that eliminate the need to internalize knowledge are also, subtly, eliminating the foundation for genuine expertise.
The Productive Struggle Problem
There’s a concept in educational research called “desirable difficulty” — the counterintuitive finding that learning conditions that feel harder and produce more errors in the short term often produce better long-term retention and transfer. Struggling with a problem — being confused, trying approaches that don’t work, getting stuck and persisting — produces deeper learning than being guided smoothly to the correct answer.
AI tools, optimized as they are for producing helpful, accurate responses quickly, are excellent desirable-difficulty eliminators. The student who would have spent 40 minutes wrestling with an algebra problem, making mistakes, erasing, trying a different approach — and in that struggle building genuine mathematical intuition — can now get a worked solution in 15 seconds.
This is not the same as the teacher showing how to solve a similar problem and then having the student practice. The student practicing — making the errors, self-correcting, developing intuition about which approach to use when — is doing the learning. The AI producing the correct answer on request is bypassing it.
What Should Actually Change
The response to AI in education has been almost entirely reactive: how do we detect AI use, how do we prevent it, how do we update our policies? These are the wrong questions.
The right questions are:
What learning activities genuinely cannot be replaced by AI? Discussion, debate, in-person oral explanation, laboratory work, physical demonstration, and collaborative problem-solving are all harder to offload to AI than written assignments. Redesigning assessment toward these modes makes pedagogical and practical sense.
How do we teach students to use AI as a tool rather than a substitute for thinking? A student who uses AI to brainstorm possible arguments and then chooses and develops one is using AI productively. A student who uses AI to do their intellectual work entirely is not developing. The distinction is teachable, and teaching it is more productive than banning everything.
What does genuine AI-assisted learning look like? Socratic dialogue with AI tutors has genuine educational potential. Using AI to get immediate feedback on drafts, identify logical gaps, and suggest resources — while still doing the thinking and writing — is meaningfully different from using AI to generate work. There’s a real pedagogy of AI collaboration to be developed.
What should be explicitly, deliberately, AI-free? Not everything, but the activities most directly connected to building foundational cognitive skills — early writing, mathematical computation, close reading of primary sources, oral argumentation — probably should be protected spaces where students develop the underlying capability before they have access to tools that can do it for them.
The Equity Problem Nobody Is Talking About
There is a class dimension to the AI-in-education problem that deserves explicit attention.
Affluent families with engaged, educated parents have the resources to ensure their children develop genuine competence, even in an AI-saturated world. Private schools with small classes, strong teacher relationships, and engaged parental monitoring can distinguish AI-generated work from genuine learning and address it.
Students in underfunded public schools, with overwhelmed teachers managing large classes and too little time for individual assessment, are more likely to have AI-assisted corner-cutting go unaddressed — and less likely to have parents at home who notice if their child has stopped genuinely learning to write, reason, and think.
The danger is that AI in education accelerates an existing bifurcation: a class of students who learn to use AI as a powerful tool while building genuine underlying expertise, and a larger class of students for whom AI becomes a substitute for developing cognitive capabilities they never fully acquire.
That gap — if it emerges — will show up in the workforce, in earnings, and in civic life, compounding existing inequalities in ways that will be difficult to attribute and nearly impossible to reverse.
The Honest Conclusion
AI tools are not going to be uninvented, and students are not going to stop using them. The educators and policymakers who are treating this as a crisis to be managed back to normalcy are wrong — there is no returning to the pre-ChatGPT classroom.
But the honest acknowledgment that must accompany any embrace of AI in education is this: the cognitive work that seems unnecessary because AI can do it is not unnecessary. It builds the mind. The student who never learns to write with difficulty, to think through problems without immediate answers, to hold knowledge in their own head rather than outsourcing it — that student is being shortchanged in ways that will matter for decades.
The goal of education has never been to produce people who can answer questions. It’s to produce people who can think. AI makes the former effortless and the latter, if we’re not careful, optional. That’s the danger worth taking seriously.
Views expressed are opinion based on educational research and observation. Individual educational contexts and AI use cases vary significantly.
