
Why learning mathematics still matters in the AI era: John Urschel's warning about independent thinking
MIT mathematician John Urschel argues that AI can accelerate proofs without replacing conceptual understanding. His 2026 research, a Microsoft study of knowledge workers and education research help explain the tension between convenience and learning.
The real question is understanding, not merely generating the right answer
Artificial intelligence is increasingly capable of assisting research in mathematics rather than only performing simple calculations. In a November 17, 2026 MIT News interview, mathematics professor John Urschel explored what that means for the process of learning. His concern is not that tools can produce answers, but that students may confuse receiving a convincing solution with developing the reasoning needed to explain it independently.
Urschel is not arguing for removing AI from research. He has used it in his own work. The distinction is between benefiting from powerful assistance and losing the habit of investigating a difficult problem. In programming education as in mathematics, running code is not the same as understanding why it works.
How Urschel uses AI in mathematics research
Urschel described the unusual experience of working with AI on two recent mathematical projects. His MIT faculty page lists 2026 preprints involving random matrices and the behavior of Gaussian elimination. One paper with Rikhav Shah, available as arXiv 2608.19189, examines long-standing questions about growth factors and numerical stability in elimination algorithms. The research illustrates the technical setting in which human reasoning and computational assistance interact.
The interview does not establish that AI automatically proved all the results without review, nor does it claim that expert mathematicians are obsolete. Discovering a promising line of reasoning and checking every condition of a mathematical argument are different tasks.
Why struggling with a problem can be useful
A learner who tries a problem independently discovers which concepts are understood and where the gaps remain. When an AI tool immediately supplies the completed solution, that process may be bypassed. An alternative is to let the student attempt a solution before receiving a hint or an example, then explain the reasoning in their own words. The relevant design choice is how much of the intellectual work remains with the learner.
Urschel emphasizes curiosity and independent problem-solving ability as enduring educational goals. Finishing an assignment more quickly does not necessarily mean learning more. Teachers therefore need assessments that reveal reasoning and transfer, not only the quality of a final submitted artifact.
- 1Attempt the problem without assistance
- 2Use AI for a focused hint or example
- 3Explain and verify every key assumption
- 4Solve a fresh case without AI
What independent research says about AI assistance
A CHI 2025 study by Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers who described 936 AI-assisted work examples. Participants' trust in generative AI was associated with a lower reported effort for some critical-thinking activities. The evidence is self-reported and observational; it does not prove that AI use causes a permanent decline in reasoning ability.
A separate 2026 Knowledge at Wharton article discusses a study of chess learners in which unrestricted on-demand AI help was associated with weaker practice gains than more controlled assistance. The people, tasks and measurements differ from Urschel's mathematics classrooms. These findings support asking better questions about the timing and design of help, not assuming one identical effect for everyone.
Implications for educators and software teams
An instructor can ask students to document which steps were attempted alone and which received AI support. In a programming class, learners might design a solution, use AI to suggest tests, and then explain the failure cases or implement a variation without assistance. In mathematics, the ability to reconstruct an argument and identify assumptions can be as important as the final result.
The video below is a historical MIT Spectrum talk in which Urschel describes his interest in mathematics. It is not a recording of the November 2026 interview. The embedded player keeps viewing on this page, while the clearly labeled YouTube link remains available for those who want the original source.
What AI changes in the role of a human thinker
As AI becomes more useful for searching, calculating and drafting, human work may shift toward interpreting and verifying outputs. A result is much less valuable when its user cannot explain the conditions under which it is valid. This suggests that reasoning, mathematical foundations and responsible review are complementary to powerful AI systems rather than obsolete skills.
Different students and subjects may require different degrees of assistance. The MIT interview provides Urschel's perspective, the arXiv work describes the underlying research area, and the Microsoft and Wharton studies examine related questions in other populations. Their results should be kept distinct. The evidence supports careful learning design, not a universal claim that all AI assistance is harmful or that it always improves understanding.
Illustrative news photo · Unsplash / Unsplash ↗
Sources and verification
- Primary sourceMIT News — John Urschel ile yapay zekâ ve matematik röportajı ↗
- Primary sourcearXiv — Entry growth in Gaussian elimination ↗
- Technical documentJohn Urschel — MIT Mathematics öğretim üyesi araştırma sayfası ↗
- Reporting / cross-checkKnowledge at Wharton — AI assistance and learning ↗
- Reporting / cross-checkMicrosoft Research / CHI 2025 — AI ve eleştirel düşünme anketi ↗
Comments
Reader question: What is your view or technical question about this story?
No comments yet. Be the first to contribute.
Reading “Why learning mathematics still matters in the AI era: John Urschel's warning about independent ” raised this question for me: price, performance and compatibility need comparison.
Editor note: Results can vary by scale. Cost, latency and user experience should be compared together.
The learning curve and accessibility also matter for actual users.
A practical case study with load testing or a pilot would help establish the trade-offs.
✎ Write a comment