AI Can Solve Homework. But Can the Student Pass the Test?

Vibeke Fængsrud, the Founder and CEO of House of Math, knows that AI can solve homework, but can it actually help a student pass a test? This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

I failed high school math. Over the next 20 years, I earned a master’s degree in mathematics, wrote 11 books on the subject, and delivered a TEDx Talk about mathematics learning. None of that happened with artificial intelligence.

Artificial intelligence has created extraordinary opportunities for students, teachers, and parents. Complex equations can be solved in seconds, difficult concepts explained in multiple ways, and homework completed faster than ever before. But we need to ask an uncomfortable question: Are students actually learning more, or simply getting better at producing the right answers?

The Homework Paradox

Imagine a student who completes an entire assignment with AI. Every answer is correct. The essay is polished, the equations are solved, and the homework receives a perfect score. The student feels confident, the parents are relieved, and the teacher sees what appears to be progress. Then comes the exam without AI, and the struggle becomes real.

This concern is backed by research. A 2025 University of Pennsylvania study, published in the Proceedings of the National Academy of Sciences, found that students using unrestricted AI assistance performed 48% better during practice but scored 17% worse on subsequent exams without AI. A correct answer is not the same as acquired knowledge!

For teachers, homework must be redefined and followed up with a small test, if it should make sense to give homework at all. When an AI agent can do the homework, testing actual knowledge becomes imperative to make sure the time is used to develop the student.

Parents face a similar dilemma. Nobody wants to watch their child struggle for hours over homework, whether it’s mathematics, writing, or science. AI can provide welcome support, explain unfamiliar concepts, and reduce tension at home. But when does helpful assistance become dependency?

The problem isn’t that AI can provide answers. It’s what happens when providing answers becomes the technology’s sole purpose.

A 2025 MIT Media Lab study used EEG to compare three groups: people writing essays with ChatGPT, people who think first, then use ChatGPT, and people who would use only their own brains. Brain activity dropped as tool support increased, while the ChatGPT users showed the weakest engagement and struggled to remember their own work. When the groups switched, former ChatGPT users never caught up to the unaided writers. Meanwhile, those who thought first and used AI afterward showed better memory and stronger brain activation. The lesson is this: brain first, then AI.

Mathematics Exposes What AI Cannot Afford to Ignore

Mathematics offers a particularly revealing example of this challenge because learning is cumulative. Every new concept builds on foundations established earlier. A student struggling with algebra may not actually have an algebra problem. The difficulty might originate in fractions, negative numbers, or multiplication concepts introduced years earlier. I know this from personal experience with thousands of students.

When struggling with mathematics, students often believe they simply aren’t good at it. What changed their academic journey wasn’t finding someone to give them the right answers, but discovering that they needed to rebuild the mathematical foundations they had missed. Once they understood those foundations, concepts that previously seemed impossible began to make sense. This is where AI has an opportunity to fundamentally improve education.

Rather than simply answering a student’s question, intelligent educational systems should help identify what the student doesn’t yet understand and why. What does the student already understand? Which prerequisite concepts are missing? Where did the confusion begin? And what should the student learn next? These questions are far more valuable to long-term learning than another instantly generated solution.

The Difference Between an Answer Engine and a Learning System

General-purpose AI models are remarkable at generating explanations, solving problems, and communicating information. They offer students access to assistance unimaginable just a few years ago. But the ability to explain a subject is not necessarily the same as teaching it effectively. Effective teaching requires understanding the learner, not just the material. This is especially important in mathematics, where identifying what a student doesn’t know can be as important as teaching something new.

Consider two very different AI interactions. In the first, a student asks how to solve a mathematical equation. The AI provides a detailed, accurate, step-by-step solution. In the second, the AI recognizes that the student hasn’t mastered a prerequisite concept. Instead of immediately delivering the answer, it reinforces the missing foundation, checks understanding, and progressively returns to the original problem.

Both interactions use AI. Only the second is deliberately structured to build independent mathematical understanding. The future of educational AI should not be defined solely by increasingly powerful language models, but by how effectively they combine sound pedagogy, meaningful assessment, and personalized learning.

Technology should adapt to what a student actually knows, recognizing when they’re ready to progress and when they need more practice.

AI Should Empower Teachers, Not Work Around Them

Teachers have always understood that students learn differently. The challenge is responding to those differences within the realities of a classroom. A teacher working with 25 students may be dealing with 25 different combinations of strengths, weaknesses, and knowledge gaps. This is where AI can become genuinely transformative.

Used thoughtfully, technology can identify patterns in student understanding, highlight weaknesses, recommend targeted practice, and track progress. These insights allow teachers to provide more focused instruction and intervention. But the teacher remains essential in supporting the process.

The objective should never be to remove teachers from the learning process. It should be to give them better information and more time to be a leader of learners. Parents deserve similar visibility. Beyond completed homework and test scores, they should understand whether their child is making genuine progress and where additional support is needed.

The Real Measure of AI in Education

Artificial intelligence has enormous potential to transform education, but we must evaluate it against outcomes that matter: stronger foundational knowledge, improved problem-solving ability, lasting retention, greater confidence, and independent thinking. We need owls, not parrots.

Engagement, speed, and homework completion may be useful indicators, but they are not substitutes for learning. Perhaps the most important question teachers and parents can ask is surprisingly simple: Can the student do it without AI?

If the answer is yes, technology has helped create a stronger, more capable learner. If the answer is no, we need to question how it is being used. I went from failing mathematics to dedicating my career to it after I realized my difficulties didn’t reflect my potential. They were a reflection of foundations I hadn’t yet learned.

AI now gives us an extraordinary opportunity to help more students faster by making that discovery much earlier. But the objective of education has never been to produce perfect answers. It is to develop people who can think, reason, and solve problems for themselves. The measure of educational AI should not be how many problems it can solve for a student. It should be how many problems the student eventually learns to solve without it.