Generative AI and Student Learning: Immediate Performance, Independent Problem-Solving, and the Risk of Dependence
Updated: Oct 29
AI can make an assignment easier to finish. Whether it helps us learn depends on what happens while we use it - and what we can do afterward.
You read an AI explanation, and suddenly the math problem makes sense. Every step looks reasonable. You move on, feeling relieved.
Then you try a similar problem without the chat open.
You’re stuck again.
That gap between following an explanation and producing one yourself raises a question worth asking: When AI helps us finish our homework, how do we know we’re actually learning?
I understand the appeal of getting one more assignment off the list. Between school, activities, and everything else competing for our attention, finishing can feel urgent. I’d just like the understanding to still be there tomorrow.
When Better Practice Scores Hide a Problem
A study published in PNAS in June 2025 offers a useful warning. University of Pennsylvania researchers Hamsa Bastani and colleagues studied nearly 1,000 high school mathematics students at one school in Turkey. Classes were assigned to practice with ordinary learning resources, a basic GPT-4 assistant, or a specially designed AI tutor.
Students using the basic assistant performed better during practice. But on subsequent exams without assistance, their scores were approximately 17% lower relative to the comparison group’s scores. That was a difference in test performance -not a 17-percentage-point drop or a decrease in intelligence.
The specially designed tutor largely avoided this decline, although it did not produce a statistically significant improvement in independent exam performance over the comparison group. It had instructions to offer hints, plus teacher-provided solutions and guidance about common mistakes.
These were short-term assessments following practice, not measurements of what students remembered months later. Still, the finding makes an important distinction: producing correct answers with assistance is different from developing the ability to solve problems independently. Read the study.
AI Can Also Support Learning
The evidence is not simply an argument against AI.
In a 2025 study published in Scientific Reports, Harvard researchers Greg Kestin, Kelly Miller, and colleagues compared a custom AI tutor with in-class active learning in college physics. Active learning involves students working through ideas rather than only listening to a lecture.
Across two lessons, students using the tutor showed greater learning gains on post-lesson tests. Their median time spent learning was 49 minutes, compared with an assumed 60 - minute classroom learning period. Not everyone finished faster: 30% of AI participants spent more than 60 minutes.
This was a carefully designed tutor tested on particular topics. The study did not include a delayed test, so it cannot say how much students remembered a week or a semester later. It does not establish that any chatbot can outperform any teacher, or that the advantage would persist across a semester. It does show why the design of the learning experience deserves attention. Read the study.
Together, these studies suggest a more useful question than “Is AI good or bad for homework?” We should ask: What thinking does the tool help us practice, and what thinking does it do for us?
Which Part of the Work Are We Handing Over?
One useful concept is cognitive offloading: using something outside ourselves to handle part of a mental task. A calendar holds appointments. A written list keeps track of groceries.
Offloading can be helpful. With schoolwork, however, the value depends partly on which skill we are supposed to learn.
If I’m learning to build an argument and AI writes the entire paragraph, I lose an opportunity to choose a claim, weigh evidence, and connect my ideas. If I draft the paragraph first and ask which claim needs stronger support, the feedback gives me something to investigate.
Consider these different uses:
Assignment | AI takes over the target skill | AI helps you practice it |
Mathematics | Generates a solution to copy | Comments on your attempted steps and offers a hint |
Writing | Produces the argument and evidence | Questions an unsupported claim in your draft |
Coding | Replaces your program with working code | Helps you investigate why a test failed |
Biology | Writes your explanation of a process | Questions your explanation so you can find gaps |
Language learning | Completes the entire exercise | Gives feedback after you attempt a sentence |
These are suggested approaches, not guarantees. An AI-generated hint or correction can still be wrong.
The aim is to keep ourselves involved in the work that builds the skill.
Confidence Is Not a Learning Test
Another useful concept is metacognition: paying attention to what we understand, where we are confused, and whether our approach is working.
A clear explanation can feel convincing. The practical check is whether we can use the idea ourselves.
Adults face a related challenge. In research presented at CHI in April 2025, Hao-Ping Lee and colleagues from Carnegie Mellon University and Microsoft Research surveyed 319 knowledge workers. Greater confidence in generative AI was associated with less self-reported critical thinking.
The study relied on workers’ accounts of their experiences. It did not establish that AI causes declining intelligence, and it did not study teenagers. Its relevance here is a question about trust: how carefully do we examine an answer when we already expect the tool to be right? Read the research overview.
Several chatbots agreeing still isn’t proof. A source needs to exist, a calculation needs to work, and an explanation needs to hold up when we question it.
A Homework Routine Worth Trying
For the next seven days, choose one subject and try four steps.
First, make an attempt. Write what you know and identify where you became confused. You do not have to stay stuck indefinitely; even a partial attempt gives you something specific to ask about.
Second, ask for a hint. Try: “Here’s my attempt. Give me one hint, and wait while I try.” This is a starting point, not a recreation of the research-tested tutors.
Third, check an important step. Compare it with your textbook, class notes, a worked example, or a teacher’s explanation. If AI cites research, open the source and check whether it supports the claim.
Fourth, close the chat and try something fresh. Solve a similar problem, write a new sentence, or explain the idea in your own words. Check your work afterward. The next day, try again before reopening the explanation.
Keep a brief note of what you needed help with and what you could do independently. This is a suggested personal routine, not a scientifically validated seven-day program.
Follow your teacher’s rules about AI use, and ask when you are unsure what help is allowed. If you remain confused, bring your attempt to a teacher or tutor. Needing another explanation does not mean you are bad at the subject.
What Should Remain After the Chat Closes?
A Bible verse that matters to me says, “Prove all things; hold fast that which is good” (1 Thessalonians 5:21, KJV). Its context is spiritual discernment. Personally, it also reminds me to ask questions about what I trust.
That includes answers from AI.
I’m still figuring out how to use these tools well. I appreciate the chance to ask another question when an explanation does not click. But I also want to notice when a finished assignment is making me feel more prepared than I actually am.
The question I want to carry into my next homework session is simple:
What will I be able to do when this chat is closed?
(Adapted from an episode of Teens Inspiring Teens: Healthy Teens, Healthy Planet, hosted by Titus Ko and produced by Teens Inspiring Teens Corporation, a 501(c)(3) nonprofit organization. Learn more at teensinspiringteens.org.)
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
The Holy Bible, King James Version. 1 Thessalonians 5:21.
