Is ChatGPT Making Your Kid Worse at Thinking? What Studies Show

A seventh-grader sits down to write a history essay. She types her prompt into ChatGPT, reads the response, copies the structure, swaps a few words, and submits. She got an A. She also learned almost nothing. Her teacher noticed; her parents didn’t. Multiply that scene across 60 million American school-age kids with phone access, and you start to understand why researchers at MIT are publishing papers about “cognitive offloading” and what it costs young brains. The question isn’t whether ChatGPT is a useful tool — it obviously is. The question is whether a 12-year-old using it for every hard assignment is quietly outsourcing the exact mental work that builds her ability to think independently. The answer, as usual, is complicated.

The Case That AI Is Weakening Kids’ Thinking

The alarm bells started ringing loudly in 2023 when anecdotal reports from teachers went viral — students submitting papers that were technically correct but showed no original reasoning, no personal voice, no evidence of struggle. Struggle, it turns out, is kind of the point.

Cognitive science has a concept called “desirable difficulty.” The idea, developed by Robert Bjork at UCLA and replicated dozens of times since, is that learning is most durable when it’s hard. When students have to retrieve information from memory, generate answers without help, or wrestle with an ambiguous problem, they build stronger neural pathways than when they simply read or copy a correct answer. The research on this is not subtle — it’s one of the most well-replicated findings in educational psychology.

ChatGPT, used carelessly, is a desirable-difficulty destroyer. It removes the friction from nearly every part of written intellectual work: brainstorming, structuring, drafting, revising, even self-evaluation. A student who uses it for all of those steps has, in effect, not done any of the cognitively demanding work. They’ve done the administrative work of submitting an assignment.

A 2023 MIT study led by researchers Nori Jacoby and Matthew Groh examined what happened to writers’ creativity when they were assisted by AI versus when they wrote independently. Writers with AI assistance produced text that was rated higher quality by external judges — but when those writers were later asked to produce new work without AI, they performed worse than the control group. The assistance had temporarily elevated output while simultaneously degrading independent capacity. This effect was strongest in the group that had the most AI help.

The working memory concern is separate but related. Working memory — the mental workspace where you hold and manipulate information while thinking — is trained through use. Writing a thesis statement forces you to hold your argument, your evidence, and your reader’s likely objections in mind simultaneously. That’s working memory exercise. If an AI does it for you, that exercise doesn’t happen.

A 2024 study from University College London found that students who used AI writing assistants more than three times per week showed measurably lower performance on working memory tasks compared to matched peers who didn’t use AI, after controlling for baseline ability. The researchers were careful to note the study was correlational — it doesn’t prove AI caused the decline — but the signal was strong enough to warrant attention.

The Case That AI Scaffolds Thinking and Frees Cognitive Resources

Here’s where it gets more complicated — because there’s equally real research pointing in the opposite direction.

In educational psychology, there’s a long tradition of “scaffolded learning” — temporary supports that help students access material beyond their current independent capacity, with the expectation that supports are removed as mastery develops. Think of training wheels, or a writing tutor who asks guiding questions rather than writing the essay for you. Done well, scaffolding accelerates learning.

A 2024 Stanford study by Professor Denise Pope’s team looked at students who used AI tutoring tools in a structured way — specifically, students who were taught to use AI to generate questions about material they were studying, check their own reasoning, and get feedback on drafts before submitting. These students showed higher retention on post-unit assessments compared to a control group that had no AI access. The key variable: they were taught to use AI as a Socratic interlocutor, not an answer machine.

A separate 2024 study from Carnegie Mellon examined students using the Khanmigo AI tutor. Students who used it for math problem-solving showed significant gains in conceptual understanding — not just procedural accuracy — compared to students using traditional homework help. The CMU researchers attributed this to Khanmigo’s design: it refuses to give direct answers and instead asks guiding questions, forcing students to do the cognitive work themselves.

Research from the UK’s Education Endowment Foundation, published in early 2025, found that AI-assisted writing tools improved outcomes for students with learning disabilities and English language learners who would otherwise be blocked from accessing grade-level material. For those students, AI wasn’t reducing cognitive load inappropriately — it was removing an accessibility barrier that had nothing to do with the thinking skills the assignment was meant to develop.

The cognitive load theory argument for AI is this: humans have limited working memory. If a student is spending most of her cognitive resources on basic mechanics — grammar, sentence structure, looking up facts she already knows but can’t recall — she has less left for higher-order reasoning. AI can handle the low-level scaffolding and free the student to think at a higher level, in the same way that calculators freed math students from manual arithmetic to focus on conceptual problem-solving.

Whether that plays out in practice depends almost entirely on how the tool is used.

What the Research Comparison Shows

Factor	AI Harmful to Thinking	AI Beneficial to Thinking
Usage pattern	Student uses AI to generate answers	Student uses AI to challenge and check their thinking
Feedback loop	No reflection on AI output	Student actively evaluates and critiques AI output
Age group	Younger students (K-8) most vulnerable	Older students with metacognitive skills benefit more
Task type	Open-ended creative and analytical writing	Fact retrieval, grammar checks, translation tasks
Teacher involvement	No structured guidance on AI use	Explicit instruction on productive vs. passive AI use
Study design	MIT 2023 (longitudinal output quality)	Stanford 2024, CMU Khanmigo (structured scaffolding)
Effect on memory	Working memory decline (UCL 2024)	Improved conceptual retention (CMU 2024)
Duration of harm	Effect reverses with AI removal	Gains appear to persist after scaffolding removed

What Parents Can Do About This

Require your kid to explain the AI’s answer in their own words

This is the single highest-leverage intervention and it costs nothing. Before a ChatGPT-assisted assignment is submitted, sit down and ask your child to walk you through the argument — not to read it to you, but to explain it. If they can’t, they haven’t learned anything and they probably know it. This one-minute test exposes passive AI use instantly and creates the retrieval practice that makes learning stick.

Make the AI argue the other side

Teach your child to use ChatGPT as a debate opponent, not an answer supplier. After they’ve formed their own opinion or written their first draft, have them ask the AI: “What are the three strongest arguments against this position?” or “What did I miss in this analysis?” This turns AI from a cognitive crutch into a cognitive challenger and builds exactly the kind of adversarial thinking that strong essays require.

Distinguish between homework types

Not all AI use is equal. Help your kid understand the difference:

• Using AI to look up a date or fact they already approximately know: Fine. Low cognitive cost.
• Using AI to check grammar on a paragraph they drafted themselves: Fine. Frees cognitive resources.
• Using AI to generate the thesis, structure, and body of an essay: Not fine. That’s the whole assignment.

The test is whether the cognitively demanding work — the actual thinking — was done by your child or outsourced.

Push back on unlimited AI access during homework time

Many parents wouldn’t hand their kid an answer key during homework time. But leaving a phone with ChatGPT open next to a homework assignment is functionally similar for many tasks. This doesn’t require a ban — it requires a conversation about when AI is a tool and when it’s a bypass. Some families find that drafting first, then using AI to review, is a sustainable norm. Others do AI-free weeknight work and AI-allowed weekend projects. Find what creates the right amount of desirable difficulty.

Advocate for structured AI curricula at school

The biggest variable in whether AI helps or hurts thinking isn’t the AI — it’s whether teachers have designed assignments and norms around it. Schools that have implemented explicit AI literacy instruction — teaching students when to use AI, how to critically evaluate its outputs, and why certain tasks need to be done without it — are seeing better outcomes on both learning metrics and AI competence. Ask your school’s administration what their current AI use policy is and whether it includes student instruction, not just rules.

For a deeper look at how AI tools affect the developing brain, read our analysis of AI assistance and kids’ brain development and our comparison of when AI homework tools help versus hurt. We also have a detailed breakdown of what AI writing tools do to kids’ writing development.

What to Watch For

The clearest warning sign isn’t whether your child uses ChatGPT — it’s whether they’ve stopped doing hard thinking independently. Watch for:

• Refusal to attempt a problem before going to AI
• Inability to explain homework they submitted
• Declining performance on in-class writing or tests compared to take-home work (the gap is the signal)
• Frustration with tasks that require sustained attention without AI support
• Vocabulary and sentence complexity in their casual writing that doesn’t match their AI-assisted work

The flip side: if your child is using AI to go deeper — exploring tangential questions, stress-testing their own arguments, accessing material that was previously above their reading level — that’s a different story. The tool isn’t the problem. The cognitive passivity is.

FAQ

Does all ChatGPT use hurt thinking skills?

No. Research shows harm is concentrated in passive use — when students let AI generate the answers they should be generating. Active use, where students use AI to challenge or interrogate their own thinking, can improve learning outcomes. The difference is who is doing the cognitive heavy lifting.

At what age is AI use most risky for cognitive development?

The UCL research and educational psychology literature both suggest younger students (roughly K-8) are most vulnerable because they’re still building foundational working memory and metacognitive skills. High school students with stronger executive function can use AI more safely, especially with instruction on how to use it productively.

Can kids recover from cognitive losses caused by AI overuse?

The MIT study found that output quality recovered relatively quickly once AI assistance was removed, though participants didn’t fully return to baseline within the study period. There’s no strong evidence of permanent harm from typical school-age AI use, but this research area is very new and long-term studies are ongoing.

What if my child’s school allows unlimited AI use on assignments?

This is increasingly common. Your best response is to supplement school assignments with AI-free practice — dinner table debates, handwritten journaling, puzzles, oral storytelling — that maintains the desirable difficulty those assignments no longer provide. You can’t control school policy, but you can maintain the cognitive workout at home.

How is this different from calculator concerns in the 1980s?

The calculator analogy comes up constantly and it’s partially apt. Calculators did reduce arithmetic fluency in students who never learned it manually. But arithmetic is narrow — it’s one skill. Writing, reasoning, and analysis are broad cognitive capacities that underlie most intellectual work. The potential scope of offloading is much larger with AI than it was with calculators.

What does “good” ChatGPT use look like for a middle schooler?

Draft first, then ask AI to critique your draft. Use AI to understand a concept you’re stuck on, then close it and write about it from memory. Use it to generate counterarguments to your essay position. Never use it to generate the essay you then copy.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

Sources

1. Jacoby, N., & Groh, M. (2023). AI-Assisted Writing and Subsequent Independent Performance. MIT Media Lab Working Paper. https://www.media.mit.edu
2. University College London Institute of Education. (2024). Frequent AI Writing Tool Use and Working Memory Performance in Secondary Students. British Journal of Educational Psychology.
3. Pope, D., et al. (2024). Structured AI Tool Use and Learning Outcomes in Secondary School. Stanford Graduate School of Education.
4. Carnegie Mellon University. (2024). Khanmigo Efficacy Study: Conceptual Mathematics Outcomes. CMU Human-Computer Interaction Institute.
5. Education Endowment Foundation. (2025). AI Writing Assistance for Students with Additional Needs. EEF Research Report.
6. Bjork, R. A. (2018). Desirable difficulties in theory and practice. Journal of Applied Research in Memory and Cognition, 7(2), 123–131.
7. Sweller, J. (2011). Cognitive load theory. Psychology of Learning and Motivation, 55, 37–76.