Metacognitive Laziness: Better Essays, No New Knowledge

Your daughter’s history essay comes back with the best grade she has had all term. She used ChatGPT to revise it. Ask her one question about the topic, though, and she can’t answer it any better than before. That gap has a name. Metacognitive laziness is what happens when a tool does the planning, checking and judging for a learner, so the work improves while the learner’s own understanding stays where it was. A 2024 randomized study in the British Journal of Educational Technology found exactly this split: ChatGPT users improved their essays the most, yet gained no more knowledge than students who had no help at all.

The good news is that the missing part is teachable, and you can check for it at the kitchen table in about four minutes.

Key Takeaways

• In a randomized lab experiment with 117 university students, the ChatGPT group had the biggest essay-score improvement, but its knowledge gain and transfer were not significantly different from the other groups (Fan et al., 2024).
• The authors call the risk “metacognitive laziness”: handing off the evaluation, monitoring and orientation steps that make learning stick.
• A separate PNAS field experiment with nearly 1,000 high schoolers found unrestricted GPT-4 raised practice grades 48%, then left students 17% worse once the tool was removed.
• A polished product is not evidence of learning. Ask your child to explain, apply and predict without the tool.
• The answer is not “no AI.” Tools designed to make kids think, like the tutor version in the PNAS study, largely removed the harm.

What Is Metacognitive Laziness? The 2024 Essay Study

Metacognition is thinking about your own thinking: setting a goal, noticing when you are confused, checking whether an answer makes sense, and deciding what to fix. Psychologists group these moves under “self-regulated learning.” Metacognitive laziness is the habit of letting an outside agent run those moves for you.

The term comes from a paper by Yizhou Fan and colleagues, published December 10, 2024, in the British Journal of Educational Technology. The team recruited 117 university students and randomly assigned them to revise an essay with one of four kinds of support: ChatGPT, a chat with a human expert, writing-analytics tools, or no extra tool. They collected what the paper calls “multi-channel” data: learning behavior, performance and motivation.

Three results matter for parents:

1. Motivation was the same. No group ended up more intrinsically motivated than another.
2. The thinking process differed. The groups showed significantly different frequencies and sequences of self-regulated learning steps.
3. Output and learning split apart. In the authors’ words, the “ChatGPT group outperformed in the essay score improvement but their knowledge gain and transfer were not significantly different.”

The researchers wrote that tools like ChatGPT “may promote learners’ dependence on technology and potentially trigger ‘metacognitive laziness’.” Their advice to learners is concrete: engage in “evaluation, monitoring, and orientation, rather than blindly following ChatGPT’s feedback solely to complete tasks efficiently.”

Why a Better Essay Can Hide Zero Learning

Here is the mechanism, stripped down. When a student revises an essay alone, the revision forces three internal jobs. First, she has to diagnose: what is weak about this paragraph? Second, she has to retrieve: what do I know that would fix it? Third, she has to judge: is the new version better, and why? Each of those jobs pulls knowledge out of memory and reorganizes it. That retrieval and reorganizing is what builds durable learning.

A chatbot can do all three jobs in one reply. It diagnoses the weak paragraph, supplies the fix and implicitly vouches for the result. The essay improves because the model’s work is now in it. The student’s memory was barely touched.

This is not only one study’s finding. Henry Roediger and Jeffrey Karpicke showed in Psychological Science in 2006 that pulling information out of memory (taking a recall test) produced “substantially greater retention” a week later than restudying the same material, even though restudying made students more confident they would remember it. Feeling fluent and actually knowing are different signals. AI help is a powerful fluency machine.

The field evidence from high school math

The Fan study was one lab session with adults. A larger test with teenagers came from Hamsa Bastani and colleagues in PNAS (June 2025). They gave nearly 1,000 high school math students access to GPT-4 in one of two forms: a standard ChatGPT-style interface (“GPT Base”) or a tutor version prompted to protect learning (“GPT Tutor”).

With access, grades on practice problems rose 48% for GPT Base and 127% for GPT Tutor. When access was taken away, the GPT Base students “actually perform worse than those who never had access,” a 17% drop. The tutor version largely removed that harm. The authors’ description of what went wrong is blunt: without guardrails, students used GPT-4 “as a ‘crutch’.”

Confidence goes up while checking goes down

Two more studies point the same way, both with adults. Hao-Ping Lee and colleagues at Microsoft Research surveyed 319 knowledge workers for CHI 2025 and found that “higher confidence in GenAI is associated with less critical thinking.” And Daniela Fernandes and colleagues, in Computers in Human Behavior (2025), found people using AI on logic problems scored three points better than a norm group but overestimated their score by four. We cover that calibration problem in depth in why kids who use AI think they did better than they did.

What the evidence does not show

Honest limits, because they matter:

• Nobody has run the Fan study on children. The participants were university students. The PNAS study involved high schoolers, but in math, not writing.
• One session is not a school year. Fan et al. measured a single task. We don’t know whether the gap widens, closes or stays put over months.
• Lee et al. is a survey, so it shows association, not cause.
• An MIT Media Lab EEG study (Kosmyna et al., 2025) reported that essay writers using an LLM showed the weakest brain connectivity and the lowest sense of ownership, with 54 participants. It is a preprint, not peer-reviewed, and small.

The pattern across these studies is consistent, though. AI reliably improves the product. Whether it improves the person depends on who does the thinking.

How to Teach Your Kid About Metacognitive Laziness

The goal is not to scare kids off tools. It is to give them a name for the moment when the tool starts thinking for them, and a habit for catching it.

Ages 5–8: The “tell it back” game

Young kids rarely use chatbots for homework, but they meet the same trap with apps and grown-ups who answer too fast. After any “how does this work?” question that a device or adult answers, play “tell it back.” Your child explains the answer to a stuffed animal in their own words, then shows it with their hands or a drawing. If they can’t, that’s fine: “Looks like the answer went in your ears but not your brain yet. Let’s try again.” You are planting the idea that hearing an answer and knowing it are two different things.

Ages 9–12: The three-question close-the-laptop check

After any homework where AI helped, the laptop closes and your child answers three questions out loud:

1. Explain: “What is the main idea, in your own words?”
2. Apply: “Use it on a new example I make up.” (A different fraction, a different historical event.)
3. Predict: “If your teacher changed one part of the question, what would change in your answer?”

Have them rate their confidence from 1 to 5 before answering, then compare. Kids find the mismatch interesting rather than shaming, especially if you play along with your own work email.

Ages 13+: The A/B learning experiment

Teens respond to data. Suggest a two-week experiment. For one unit, they study with AI the way they normally do. For the next, they use AI only after a first attempt, and only to ask questions (“Where is my reasoning weak?”) instead of asking for fixes. Before each quiz, they predict their score. After, they log predicted versus actual in a two-column note. The Fan and Bastani studies give them a hypothesis to test on themselves. Many teens find the “AI after first attempt” units feel harder and score better, which is the point.

The question to ask: “If I took away the AI right now, what could you still do that you couldn’t do yesterday?”

Output vs. Learning: A Side-by-Side Comparison

What you can see	What it tells you	What it doesn’t tell you
A higher essay grade	The final product met the rubric	Whether your child could write it again
Fast homework completion	The task got done	Whether any retrieval or reasoning happened
Your child feels confident	Fluency was high	Whether knowledge is stored (Roediger & Karpicke found restudying raises confidence but lowers week-later recall)
They can explain it with the screen open	They can read it back	Whether it lives in memory
They can explain it with the screen closed	Real understanding, at least short term	Long-term retention (check again in a week)
They can apply it to a new example	Transfer, the outcome Fan et al. found unchanged by ChatGPT	Nothing major. This is the strongest home signal you have.

What to Do at Home When AI Is in the Homework

Ask for the draft before the polish

Make “first attempt is yours” the family rule. The child writes the rough paragraph or solves the problem badly first; AI can come in after. This keeps the diagnose-retrieve-judge cycle in the child’s head, which is the exact part the Fan study suggests gets skipped. It also mirrors the design of the GPT Tutor condition in the PNAS study, which withheld full answers.

Change the prompt from “fix it” to “find it”

“Rewrite this paragraph” hands over all three jobs. “Tell me which sentence is weakest, but don’t rewrite it” hands over only the diagnosis and leaves the fixing to your kid. Write two or three of these “find, don’t fix” prompts on a sticky note by the computer. Our guide to AI and the illusion of learning in homework has more prompt swaps.

Check transfer, not the page

The fastest home test is a new example. If your child used AI to understand photosynthesis, ask what would happen to a plant in a closet. If they used it for a persuasive essay, ask them to argue the opposite side for 30 seconds. Transfer was the outcome that did not improve in the Fan study, so it is the one worth checking.

Space the check

Ask the same “explain it” question again three to seven days later. Roediger and Karpicke’s results showed the difference between restudying and retrieval only appeared on delayed tests. A week-later check catches what a same-day check misses.

What not to do

Don’t treat a great AI-assisted grade as a lie or the tool as cheating by default. The essay may be genuinely better and your child may have followed the teacher’s rules. Interrogating every assignment teaches kids to hide AI use, not to think. Ask about learning, not about the tool. If you want more background, our explainer on teaching kids how they learn covers metacognition without the AI angle.

What to Watch For Over the Next 3 Months

• Week 4: Your child can do the close-the-laptop check on most AI-assisted assignments, even if answers are rough. They start saying things like “I don’t really get this part,” which is metacognition working.
• Month 2 red flags: Homework grades climb while quiz and test grades stay flat or drop. Your child gets irritated or anxious when asked to explain work without a screen. They can’t name a single thing they figured out themselves this week.
• Month 3 self-check: Ask your child to predict their next test score, then compare. If predictions are getting closer to reality, their self-monitoring is improving. If they keep predicting high and scoring lower, revisit the “first attempt is yours” rule.

Frequently Asked Questions

What does metacognitive laziness mean in plain English?

It means letting a tool do the “thinking about your thinking”: noticing what’s weak, deciding what to fix and checking if it’s right. The term comes from Fan and colleagues’ 2024 study, where ChatGPT users improved their essays most but learned no more than others. The work got better; the learner didn’t.

Does this mean my kid shouldn’t use ChatGPT for homework?

Not necessarily. The PNAS high school study found that a tutor-style version of GPT-4, which gave hints instead of answers, largely avoided the learning loss. The risk comes from using AI to finish tasks rather than to think through them. Rules about when and how matter more than a blanket yes or no.

Was the metacognitive laziness study done on children?

No. Fan et al. studied 117 university students in a single lab session. The closest evidence for teens is the 2025 PNAS math experiment with nearly 1,000 high schoolers, which found a similar pattern. Child-specific research on writing is still thin, so treat the findings as a strong warning sign, not a settled verdict.

How can I tell if my child learned something or AI did the work?

Close the laptop and ask them to explain the idea in their own words, apply it to a new example you make up, and predict what changes if the question changes. If they can do all three, real learning happened. If they can only explain it with the screen open, the knowledge lives in the screen.

Is a better grade with AI still a good thing?

It can be, if the teacher allowed AI and your child understands the result. But a grade measures the product, and research shows product and learning can split apart. Treat the grade as one signal and the close-the-laptop check as the other. When they disagree, believe the check.

---

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. Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., et al. (2024). “Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance.” British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
2. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., et al. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122
3. Roediger, H. L., & Karpicke, J. D. (2006). “Test-enhanced learning: Taking memory tests improves long-term retention.” Psychological Science, 17(3). https://doi.org/10.1111/j.1467-9280.2006.01693.x
4. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., et al. (2025). “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713778
5. Fernandes, D., Villa, S., Nicholls, S., Haavisto, O., Buschek, D., Schmidt, A., et al. (2025). “AI makes you smarter but none the wiser: The disconnect between performance and metacognition.” Computers in Human Behavior, 175, 108779. https://doi.org/10.1016/j.chb.2025.108779
6. Kosmyna, N., et al. (2025). “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv (preprint, not peer-reviewed). https://arxiv.org/abs/2506.08872