AI in the Classroom: Benefits, Risks, and the Impact on Learning

Kim Langen, Founder and CEO of Spirit of Math, outlines some of the benefits and risks of using AI in the classroom and explores its impact on learning. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

A student sits down with a difficult math problem, opens an AI chatbot, and has the correct answer within seconds, complete with steps that look reasonable and an explanation that reads clearly, so the homework gets finished and the technology appears to have made learning easier. Ask that same student to solve a similar problem with the chatbot closed, and you will likely see something very different.

This is one of the biggest challenges I see in education as artificial intelligence becomes part of everyday classroom life. A chatbot can produce answers, explanations, practice questions, and feedback within seconds, and every one of those can help a student get an assignment finished.

What a finished assignment shows is that a correct answer was produced somewhere, and the more important question for me is whether the student understood the material well enough to reach that answer alone. Schools investing in education technology are dealing with two separate outcomes, completing work and learning from it, and a tool that speeds up the first may do nothing for the second.

Students and teachers are already using AI, so the useful question is a practical one. Which applications help students become stronger, more independent thinkers, and which ones simply produce finished work for them?

The evidence, and what I have seen over more than 30 years of teaching, points to a clear answer. AI supports learning when it keeps the student doing the thinking. That principle applies to tutoring tools built on learning science and to classroom activities that treat AI output as something to examine. It also applies to evaluations that test what students can do on their own.

Where AI Is Showing Genuine Promise

I think AI’s most promising use is to give a student extra support when they get stuck. In a traditional classroom, a teacher may be responsible for more than 20 students, each working at a different pace. While one student needs another explanation of fractions, another is ready for a harder application, so giving each of them individual attention at exactly the right moment is hard even for very experienced educators.

Well-designed AI tutoring systems can help with part of that need by offering additional examples, answering questions, and guiding students through material outside regular classroom instruction, and there is encouraging evidence that this approach can work.

A 2025 randomized controlled study published in Scientific Reports tested an AI-powered tutor in an undergraduate physics course at Harvard University. Students who used the carefully designed tutor learned significantly more in less time than students who received the same material in an active-learning classroom session, and they also reported greater engagement and motivation (Source: Scientific Reports, 2025).

How that tutor was built made the difference. It followed established principles from education and learning science, so it asked questions, caught misunderstandings, and timed its support to each student’s needs. Those are the habits of a tool designed to teach, and a chatbot that completes a student’s work on request has none of them (Source: Scientific Reports, 2025).

AI can also help teachers identify patterns in how students perform. When many students struggle with the same type of problem, digital assessments and analytics make that weak spot much easier to find, and teachers can then decide what needs another explanation or more practice. That information helps students once a teacher knows how to act on it, which is why a teacher’s judgment still matters so much.

The Problem with Making Learning Too Easy

The speed that makes AI so useful can also work against learning. Education has always involved productive difficulty, where students meet unfamiliar ideas, test possible solutions, make mistakes, and slowly build the ability to work through challenges, and AI can remove those obstacles within moments.

Learning is hard and, at times, uncomfortable, and children need to learn that it is supposed to feel that way. When a student tries a method on a multi-step algebra problem, watches it fail, and goes back to try again, that struggle is what builds the ability to solve the next problem. A chatbot that supplies the full solution after the first attempt takes away the part of the work that does the teaching.

Research is starting to confirm this concern. The OECD’s Digital Education Outlook 2026 distinguishes between improvements in task performance and actual learning gains. Its review notes that students using general-purpose generative AI can produce higher-quality work without demonstrating better understanding once the AI is removed, and that, in some studies, the early advantage disappears or even reverses on independent assessments (Source: OECD, 2026).

Mathematics shows this very clearly, because a correct answer can hide weak reasoning. Two students can reach the same result, one with a firm grasp of the underlying concept and the other by following a sequence of instructions without knowing why those steps work.

An unfamiliar problem exposes the difference right away. A student who understands the concept can reconsider their assumptions and try another method when the numbers or the structure change, and a memorized procedure tends to break down at exactly that point.

Why Explaining an Answer Still Matters

The safeguard against bypassing the thinking is to make that thinking visible. Over more than 30 years in mathematics education, one principle has held up for me: students need opportunities to explain how they think.

Classrooms built around that principle share some habits, whatever the program or curriculum. Students work through challenging problems, compare approaches, and explain their solutions to their teachers and classmates, so that each student understands why a method works and considers whether another approach is possible.

This kind of learning feels uncomfortable at first. A student may find that a classmate solved the same problem with a completely different method, or may have to explain an assumption they never consciously examined, and those conversations are what build mathematical judgment.

AI can support this work when its role is chosen with care. A student can ask an AI tool for two different solutions to one problem and judge whether both are valid. A teacher can present an AI-generated explanation that contains a subtle mistake and ask students to find it. Students can also compare their own reasoning with an AI response and explain which approach is more efficient. In each of these activities, the AI output becomes something to examine, and the student is still the one who decides whether it is correct.

As educators, our job is to keep responsibility for the thinking with the student, so technology can supply the examples and feedback while the learner does the reasoning.

Preparing Students for a World Where AI Can Do the Work

These habits carry over into a world where AI handles more and more routine intellectual tasks. In that world, students will need to evaluate information, question assumptions, and solve problems that arrive with no obvious procedure, and those abilities only grow through practice.

When a student learns to defend a mathematical argument, they are practicing how to support a conclusion with evidence, which is a skill they will use in every reasoned argument they ever make. Additionally, when a student discovers an error in a proposed solution, they develop the judgment to question information that looks credible at first glance, and that habit becomes more valuable as AI systems produce explanations that sound convincing even when they are wrong.

When a student works through a difficult problem, they learn to keep going when the next step is uncertain. No tool can build that persistence for them, because it only comes from working through the uncertainty themselves.

These experiences keep their value no matter how sophisticated AI becomes, and the strongest educational uses of AI will recognize that by creating new opportunities for practice while keeping the learner at the center of the learning process.

How Schools Should Measure Whether AI Works

Recognizing those applications means measuring the right outcomes. Engagement figures, completion rates, and minutes saved tell educators whether students are using a product. When success is judged by the quality of completed assignments, it becomes very easy to mistake technological assistance for student ability.

I believe a better evaluation starts with what students can do independently after using the technology. Suppose an AI tutoring system helps a student understand proportional reasoning. A week later, can that student apply the concept to an unfamiliar problem, explain the relationship in their own words, and recognize a proposed solution that makes no sense? Those answers reflect something far more durable than a successful interaction with software, and the OECD’s 2026 review also emphasizes the importance of distinguishing AI-assisted performance from lasting learning gains.

The design of the Harvard tutor points to a second set of questions, about how a tool shapes student behavior over time. I would look at whether students attempt a problem before asking for help, and whether the hints they receive help them make progress without revealing the whole solution. A tool that also asks students to explain their thinking reinforces the same classroom habit I described earlier.

Fit is a separate question, and teachers have a very practical role in judging it. An AI product can be technically sophisticated and still be a poor match for a particular age group, subject, or learning objective. A tool that supports an undergraduate physics student will not automatically suit an eight-year-old learning multiplication, since those learners differ in their needs, prior knowledge, and developmental stage. Schools need evidence about students like their own, along with educators who can interpret that evidence and decide how the technology fits into their instruction.

What schools should be looking for is situations where AI helps students develop knowledge and skills they can retain and apply, and how much AI is used during the school day should follow from that. For me, the real test is what happens when the screen is closed. If the student from the opening example can now solve a similar problem without the chatbot, AI contributed to their learning. If they can only perform while the tool is open, the learning still has to happen.