AI Overconfidence in Students: Why Kids Think They Did Better

AI overconfidence in students has now been measured, and the number is specific. In a 2025 study in Computers in Human Behavior, people who used ChatGPT on law-school logic problems scored about three points better than a comparison group. But they believed they had done about four points better still. The tool made them more competent and less accurate about their competence at the same time. And the people who knew the most about AI were the worst judges of their own work.

If your child finishes an AI-assisted assignment saying “I nailed it,” this is why that feeling may not match the grade, or the next test.

Key Takeaways

• In Study 1 of Fernandes et al. (2025), 246 adults using AI on 20 LSAT logic problems improved by three points against a norm group but overestimated their performance by four points. A second study with 452 people replicated it.
• Higher AI literacy was linked to lower accuracy in judging one’s own score.
• The classic Dunning-Kruger pattern, where low scorers overestimate and high scorers are closer to accurate, disappeared when people used AI.
• Feeling confident is a weak guide to learning even without AI: in a 2006 Psychological Science study, restudying raised students’ confidence but produced worse recall a week later.
• The fix is calibration practice: predict a score, do the task without the tool, compare, repeat.

What the Research on AI Overconfidence in Students Measured

Metacognitive monitoring is the ability to judge how well you are doing while you do it. It’s what lets a student say “I think I got 14 of 20 right” and be close. Psychologists measure it by comparing what people predict with what they score.

Daniela Fernandes and colleagues published “AI makes you smarter but none the wiser” in Computers in Human Behavior on October 9, 2025. In Study 1, 246 participants used an AI system to solve 20 logical reasoning problems from the Law School Admission Test. Then they estimated how many they got right.

The headline results, from the paper’s abstract:

• “Task performance improved by three points compared to a norm population.”
• “Participants overestimated their task performance by four points.”
• “Higher AI literacy correlated with lower metacognitive accuracy, suggesting that those with more technical knowledge of AI were more confident but less precise in judging their own performance.”
• The Dunning-Kruger effect, “usually observed in this task, ceased to exist with AI use.”
• Study 2, with 452 participants, replicated the findings.

So the AI did help. The problem is the gap between helped and felt helped.

Why AI Inflates the Feeling of Competence

Here is the mechanism in engineering terms. Your brain estimates performance from cues it can sense: how fluent the work felt, how quickly answers came, how confident the reasoning sounded. Without a tool, those cues come from your own processing, so they correlate, roughly, with your skill.

With a chatbot, the cues are borrowed. The answer arrives fluently because a language model generated it fluently. The reasoning sounds confident because models are trained to produce confident, well-formed text. Your brain reads those cues as “I understood this,” but the signal came from the machine’s output, not from your understanding. Think of it as a sensor reading the wrong source: the gauge goes up whether or not your own engine is running.

That also explains the odd AI-literacy result. People who use AI a lot may trust its output more and check it less. A Microsoft Research survey of 319 knowledge workers (Lee et al., CHI 2025) found a similar pattern from a different angle: “higher confidence in GenAI is associated with less critical thinking.”

What happened to Dunning-Kruger

In 1999, Justin Kruger and David Dunning published “Unskilled and unaware of it” in the Journal of Personality and Social Psychology. Across four studies, people in the bottom quartile on tests of humor, grammar and logic scored around the 12th percentile but estimated themselves at the 62nd. Improving their skills helped them see their limits.

Fernandes and colleagues report that this pattern vanished with AI. The authors’ interpretation is that AI “levels cognitive and metacognitive performance.” That sounds like good news, but the overall result was still overestimation. The AI didn’t make people well calibrated. It made the low and high scorers wrong in a more similar way. We go deeper on the original effect in the Dunning-Kruger effect in kids.

The feeling of learning has always been unreliable

AI didn’t invent this problem; it amplified one learning scientists already knew. Henry Roediger and Jeffrey Karpicke (Psychological Science, 2006) had students either restudy prose passages or take recall tests. On a test one week later, the recall-test group remembered substantially more, “even though repeated studying increased students’ confidence in their ability to remember the material.”

And in a 2019 PNAS study, Louis Deslauriers and colleagues found physics students in active-learning classes learned more but felt they learned less than peers in lectures. The extra effort felt like failure. Our piece on why effortful learning feels worse and works better covers that study in full.

Put together: easy, fluent work makes everyone feel more competent. AI is the most fluent study partner ever built.

The limits of this evidence

Fernandes and colleagues studied adults on logic puzzles, not kids on homework. No published study has measured this exact gap in children, as far as we could find. It’s reasonable to expect children’s self-monitoring to be weaker than adults’ to begin with, but that is an inference, not a finding. The study measured a single sitting, so we don’t know whether calibration improves as people get used to AI. And the “norm population” comparison means AI users were compared to a reference group, not to themselves without AI.

The closest evidence with teens is the 2025 PNAS field experiment by Hamsa Bastani and colleagues: nearly 1,000 high school math students with unrestricted GPT-4 raised practice grades by 48%, then scored 17% worse than peers once access was removed. Practice performance was a poor predictor of real skill.

Confidence Signals Compared: What to Trust

Signal your child reports	How reliable it is with AI help	Better check
”That was easy”	Low. Ease often comes from the tool’s fluency	Ask them to redo one problem cold
”I’m sure I got it right”	Low. Fernandes et al. found a 4-point overestimate	Have them predict a score, then grade it
”I understand it now”	Medium. Depends on whether they can say it without the screen	Close-the-laptop explanation
”I could teach it to my friend”	Medium to high if they actually do it	Let them teach you for two minutes
Correct answers on a new, similar problem with no AI	High	This is the real test
Steady grades on in-class quizzes	High	Compare with homework grades over a month

How to Calibrate a Child’s Sense of What They Know

Make “predict, then check” a habit

Before any quiz, worksheet or practice set, your child writes a number: “I think I’ll get 8 out of 10.” After it’s graded, they write the real number next to it. That’s it. Over weeks, the gap usually shrinks, because the brain starts using better cues. Keep it neutral. A gap is information, not a failure.

Separate “with tool” and “without tool” scores

If AI is allowed for homework, ask your child to do the last one or two problems without it. Compare. If they get 10 of 10 with help and 3 of 5 without, that tells both of you where they actually are. For teens, a simple spreadsheet with three columns (predicted, with AI, without AI) turns this into an experiment they own.

Ask for the reason, not the answer

Overconfidence hides behind correct answers. “Why is that right?” or “What would make it wrong?” forces your child to generate reasoning, which is the cue their brain should be using to judge competence. If the reason is “the AI said so,” they’ve found the gap themselves.

Let them be wrong in low-stakes places

Calibration needs feedback, and feedback needs mistakes. A kitchen-table practice quiz with no grade attached is a better calibration tool than a high-stakes test, because your child can afford to discover they were overconfident.

What not to do

Don’t tell your child “you’re not as smart as you think.” The research shows overconfidence with AI is a general human pattern, including among AI-savvy adults. Framing it as a flaw in your kid makes them defensive. Framing it as a known glitch in how everyone’s brain reads fluency makes it something to test, which kids tend to enjoy.

What to Watch For Over the Next 3 Months

• Week 4: Your child has a few weeks of predicted-versus-actual scores. Look at the pattern, not any single number. Are predictions consistently high?
• Month 2 red flags: A widening gap between homework grades and test grades. Surprise and frustration after tests (“but I knew this!”). Refusing to try problems without the tool.
• Month 3 self-check: Predictions are within a point or two of actual scores most of the time. Your child can name one topic they thought they knew but didn’t. That kind of self-correction is the skill you’re building.

Frequently Asked Questions

Does AI make students overconfident?

The best current evidence says yes, at least in adults. Fernandes et al. (2025) found people using AI on logic problems overestimated their scores by about four points, and a second study replicated the result. No published study has yet measured the same gap in children, so treat it as a likely pattern, not a proven one.

Why would knowing more about AI make my teen worse at judging their work?

The study found that higher AI literacy correlated with lower accuracy, and the authors suggest technically knowledgeable users were more confident but less precise. One plausible reason: frequent users trust the output more and check it less. It’s a correlation, so it doesn’t prove that learning about AI causes overconfidence.

Is this just the Dunning-Kruger effect?

Not quite. Dunning-Kruger describes low performers overestimating the most. With AI, Fernandes and colleagues found that pattern disappeared, yet people still overestimated overall. It’s a different shape of error: more uniform, and tied to the tool rather than to low skill.

How can I tell if my child really knows the material?

Have them solve a new, similar problem without AI and explain why the answer is right. Ask them to predict their score first. If they solve it, explain it, and predict close to the result, they know it. A correct AI-assisted answer alone doesn’t tell you much.

Should I ban AI for homework because of this?

The research points to how AI is used more than whether. Overconfidence is a monitoring problem, and monitoring can be trained with prediction and no-tool checks. Many families keep AI for some tasks and add a short “without the tool” step at the end. Follow your school’s rules first.

---

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. 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
2. Kruger, J., & Dunning, D. (1999). “Unskilled and unaware of it: How difficulties in recognizing one’s own incompetence lead to inflated self-assessments.” Journal of Personality and Social Psychology, 77(6). https://doi.org/10.1037/0022-3514.77.6.1121
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. Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. (2019). “Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom.” Proceedings of the National Academy of Sciences, 116. https://doi.org/10.1073/pnas.1821936116
5. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., et al. (2025). “The Impact of Generative AI on Critical Thinking.” CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713778
6. 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