Why AI should not be used in education is a fair question, and the honest answer is not a flat yes or no. A 2025 Microsoft and Carnegie Mellon study of 319 knowledge workers found that higher confidence in an AI's answer predicted less critical thinking, while higher confidence in one's own judgment predicted more. A separate MIT Media Lab study of 54 students found up to 55 percent less brain connectivity and higher cognitive load among those who used an AI chatbot for essay writing than those who worked unaided, a pattern researchers called accumulating cognitive debt. Teachers report a related cost: once a submission is suspected of being AI-written, trust between teacher and student breaks down. The fix the evidence supports is not a ban, it is answering first, then asking, then comparing the two.
On this page+
Ask a teacher why they are wary of AI in the classroom and you rarely get a rant about robots. You get a smaller complaint. A student who used to struggle, then think, then write, now hands in something clean on the first try. They cannot explain how they got there. Whether that is a problem, and how big a one, is what "why AI should not be used in education" is really asking.
Searches for AI overreliance in education and does AI harm critical thinking are really asking the same thing. The honest answer is not a flat yes or no. Two recent studies, a set of teacher accounts, and one large faculty survey point at the same mechanism from different angles: it is not the AI that erodes thinking, it is the specific act of trusting its answer instead of building your own. That distinction changes what you do next, and it is the one most of the "pros and cons" lists skip.
Why AI should not be used in education, according to two 2025 studies
The clearest evidence says: it depends on how much you trust the tool, not whether you touch it. A 2025 study by Lee and colleagues at Microsoft Research and Carnegie Mellon, presented at CHI 2025, surveyed 319 knowledge workers about 936 real instances of using generative AI at work1. The pattern held across almost every task: the more confidence someone placed in the AI's answer, the less critical thinking they reported doing to check it. The reverse also held. People with higher confidence in their own judgment kept thinking critically even while using the tool. Confidence in the machine and confidence in yourself pulled in opposite directions.
A separate MIT Media Lab study adds a second angle: what repeated use looks like in the brain, not just in a survey2. Researchers led by Nataliya Kosmyna had 54 participants write essays across three conditions: using an AI chatbot, using a search engine, or writing unaided, across four sessions. The AI-chatbot group showed up to 55 percent less brain connectivity than the unaided group, plus measurably reduced attention, a pattern the researchers called accumulating "cognitive debt." Together, the two studies say something specific: not "AI makes you dumber," but "trusting an answer removes the friction critical thinking runs on."
What happens when a student hands over the thinking, not just the task?
A high school English teacher writing in Education Week put it plainly after tracking more than 100 AI-generated submissions passed off as original work since 2022: "education is about the process of learning, not the product." That is worth sitting with, because it names what a shortcut actually shortcuts. A finished essay is not the point of assigning one. The point is what happens to the student while they build it: the false starts, the sentence they rewrote four times, the argument that only became clear once they tried to defend it on paper. Cognitive offloading, handing a mental task to a tool instead of doing it yourself, is a perfectly normal thing to do with a calculator or a spreadsheet, because arithmetic was never the skill being taught. Handing off the reasoning in a reasoning class is a different trade, because the reasoning was the assignment.
This is also where the "why not" case is strongest and most specific: not that AI exists, but that certain assignments were designed to build a skill through friction, and removing the friction removes the only place that skill was going to form.
The evidence does not say AI destroys thinking. It says trusting the AI's answer is what stops you doing your own.
Why do teachers say AI erodes trust in the classroom?
The practical fallout shows up before the research does. A Harvard Independent op-ed from a current student described watching classmates submit AI writing as their own, and cited Dean Martin West's description of the tool doing "the cognitive work of thinking for students" rather than supporting it3. The same piece caught an early version of a popular AI chatbot fabricating a statistic and inventing a source to support it, a reminder that the tool's confidence in its own answer is not evidence the answer is right.
What both teacher and student accounts describe next is a relationship cost that does not show up in a pros-and-cons list: once a teacher suspects AI use, every subsequent submission gets read with suspicion first and engaged with second. That is expensive in a way that has nothing to do with test scores. A classroom runs on the assumption that the work in front of a teacher is a window into the student's thinking. When that assumption breaks, the teacher stops being able to see the thing they are actually there to help with.
Can AI make some students better thinkers instead of worse ones?
Yes, and the research on AI and critical thinking keeps landing in the same place: the same CHI 2025 study is the evidence for that too. It was not confidence in AI alone that predicted less critical thinking. It was confidence in AI relative to confidence in yourself. A student who treats the AI's draft as a first guess to interrogate, rather than a finished answer to submit, is doing something closer to using a research assistant than outsourcing a task. The tool did not decide which one happened. The user's own posture toward the answer did.
This is a genuinely uncomfortable finding for a simple "ban it" policy, because it means the same tool produces opposite outcomes depending on something a school cannot easily police: whether the student who opened it was already confident enough in their own judgment to argue with what came back. It is also why a tool built to slow that moment down, rather than resolve it, changes the outcome: Inwitt's own Critical thinking mode exists for exactly this gap, asking the follow-up question instead of supplying the conclusion.
Using AI to answer
Using AI to think
You ask, then submit what comes back
You answer first, then ask, then compare
The gap between your guess and the AI's version disappears unexamined
The gap is the part you learn from
Confidence sits with the tool
Confidence stays with you, and grows with practice
A skipped semester of practice, invisibly
A semester of practice, with a shortcut for the boring parts
What does the evidence not yet tell us?
Some honesty matters here, because overclaiming what the research shows is its own kind of harm. The CHI 2025 study is self-reported and correlational: it shows confidence and critical-thinking effort moving in opposite directions, not that one directly causes the other to drop. The MIT study is also small, 54 participants, measuring attention and connectivity over four sessions, not learning outcomes over a semester or a degree.
A national survey of 1,057 faculty by the American Association of Colleges and Universities and Elon University, conducted in late 2025, found that 95 percent expect AI to increase student overreliance and 90 percent expect it to diminish critical thinking skills4. The survey's own authors call it non-scientific and not generalizable. An expectation is not a measurement of what happened in a classroom. None of this proves AI cannot be taught around. It proves the risk is real enough that "just let them use it" and "just ban it" are both too simple to be honest answers.
Common mistakes schools and parents make
AI dependence in the classroom is not something a blanket rule fixes. Banning the tool outright treats the object as the problem, when the CHI 2025 finding says the posture toward the object is what matters; a ban also does nothing for the identical offloading habit a student can practise on a search engine or a friend's answer key. Assuming every AI-assisted submission is cheating skips the difference between "I asked it to write my argument" and "I asked it to check an argument I already wrote," a difference a rubric can actually test for. And ignoring the self-confidence finding is the biggest miss: a policy built only around detection will always be a step behind the tool, while teaching AI literacy alongside a habit of answering first is not.
Frequently asked questions
Does AI in education help or hurt critical thinking?
Both, depending on use. The research shows it is confidence in the AI's answer relative to confidence in your own judgment that predicts the effect, not the tool's mere presence.
Is it fair to blame AI for declining critical thinking skills?
Partly. Faculty surveys and small studies show a real association, but the evidence is correlational and the sample sizes are modest. Skills that were already weak before AI existed do not become AI's fault by association.
How can a student use AI without losing the skill?
Write your own answer or approach first, then ask the AI, then compare the two. Treat what comes back as a draft to argue with, not a submission to copy.
Should schools ban AI entirely?
The evidence does not support that as the fix. It supports teaching the habit of answering first, since that is the variable the research actually ties to the outcome.
Most advice here tells a student what to conclude about AI. A genuinely useful thinking partner would ask you to notice, in your own work this week, the moment you reached for an answer instead of building one, since that moment is the one the research keeps pointing back to. That is the premise behind Inwitt AI, a thinking partner app built around asking one sharp question at a time instead of answering: its research mode locks findings behind a guess-first reveal for exactly this reason, so you commit to your own answer before you see anyone else's, AI included. If you want to try the habit this piece describes on a real decision or assignment, your first session walks through what that looks like, and the waitlist is open now.
Footnotes
Footnotes
Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson, The Impact of Generative AI on Critical Thinking, CHI 2025 (Microsoft Research and Carnegie Mellon University), 2025. ↩
