Artificial intelligence can help solve mathematical problems quickly, but getting an answer is not the same as understanding it. John Urschel, a mathematics professor at MIT, argues that students still need to develop their own problem-solving skills and curiosity, even as AI becomes part of education.
Learning math means working through uncertainty
For Urschel, the appeal of mathematics begins with puzzles. The process is to face a problem that is not yet clear, find a way into it and gradually reach understanding. That experience, he says, is rewarding in its own right.
He wants students to strengthen their ability to think quantitatively and solve problems. Learning the central ideas and techniques in a particular course matters, too: those fundamentals give students tools they can use to tackle problems beyond that class.
This view puts the work of learning at the center. A student does not build problem-solving ability simply by seeing a finished solution. The effort of finding a path through a problem is part of what develops that ability.
AI can speed up mathematical work
Urschel recently published two preprints on arXiv, one on the growth factor of random matrices and the other on worst-case behavior in Gaussian elimination. He used human insight alongside AI tools to address problems that had stumped mathematicians for decades.
He describes using AI in this work as something like having a collaborator with access to extensive resources that can complete certain tasks much faster than he can. In his view, AI systems can already do some kinds of mathematics more efficiently than people.
That speed changes what assistance can offer, but it does not remove the value of human understanding. If a system produces a proof, Urschel says, people need to understand how its mathematical ideas fit together, absorb that reasoning and be able to explain it to others.
Getting an answer is not the same as learning
Urschel finds solving a problem on his own more joyful than doing so with AI. He sees independent problem-solving ability and curiosity as important skills, including for students who use AI tools.
He puts understanding at the heart of that use. A generated proof may be a useful starting point, but the learner still has to make sense of the argument. Without that step, a polished result can create a misleading sense of competence: someone may feel they understand a topic when they have not learned it deeply.
The same concern applies to homework and writing. AI can make it easy to produce work quickly, especially when students are tired or busy. But finishing an assignment with little engagement may bypass the practice that was meant to help them learn.
Access and effort both matter
Urschel also sees technology as an opportunity. He recalls using a laptop and Google to find advanced mathematical concepts and papers by leading experts while he was in high school and college. That access helped his education, and he sees AI as extending what learners can reach and how quickly they can explore it.
At the same time, easier access can make it tempting to hand over the thinking. Urschel argues that students need to engage with material and push themselves through difficulty. The struggle of learning can feel uncomfortable, but skipping it may leave students with the impression of learning without the understanding.
That places responsibility on instructors as well as learners. Urschel points to useful recitations and office hours, and to encouraging students to take responsibility for their education by showing up and engaging with the material. In an AI-enabled classroom, assistance can help students move forward; the lasting value depends on whether they understand what they have learned.