Smartest AI vs Professor-Made Math Problems: Student Preferences Reveal the Winner

KOMPAS.com – Fears that artificial intelligence (AI) will replace the role of mathematicians were flatly dismissed by Martin Hairer, winner of the 2014 Fields Medal (the most prestigious award in the world of mathematics).

According to Hairer, mathematics still “safe” from threats AI.

Hairer expressed this when responding to the anxiety of a high school student who sent him an e-mail.

The student is worried that his future in mathematics will be lost because AI capabilities are becoming more sophisticated every day.

Hairer calmed the student with strong arguments. He admitted that Large Language Models (LLM) such as ChatGPT are indeed good at solving standard practice questions that already have answers on the internet.

However, he emphasized that he had never seen a single example where AI was able to generate truly original mathematical ideas or concepts.

“I have yet to see a reasonable example where an LLM has resulted in a truly original new idea or concept,” Hairer said.

To prove this claim, Hairer and a team of elite mathematicians from Harvard, Stanford, and MathSci.ai released an experiment entitled “First Proof”.

In this experiment, they tested today’s most advanced AI models, including ChatGPT-5.2 Pro and Google Gemini 3.0 Deep Think, using previously unpublished research problems.

Also read: Moltbook, Medsos Where AI Chats and Talks to Humans

The goal is clear, ensuring that AI cannot “cheat” answers from its training data on the internet.

The result? Disappointing. Hairer even compared the AI’s answers to that of a “less intelligent undergraduate student.”

“AI models tend to provide a lot of long-winded details in the easy parts. However, at the heart of the difficult arguments, the details are minimal,” Hairer said.

He added that AI seemed to know the starting point and final destination, but did not know the way to get there.

As a result, AI often just “makes up” or does hand-waving (giving unclear/vague arguments) in the middle of the proof, hoping the reader won’t notice.

Also read: Graduates from famous campuses are no longer safe in the AI era

Weaknesses of AI

The research team found several reasons why AI still cannot replace mathematics professors:

• Weak at visuals, AI is notoriously bad at visual reasoning (visual reasoning), so questions that require spatial imagination are bound to fail.
• Short memory, if the answer requires proof that is more than 5 pages long, the quality of the AI answer will immediately drop and it will start to be inconsequential.
• Just “Yes Man”, Tamara Kolda, one of the authors of the paper, called AI boring because it couldn’t be debated. AI only follows the user’s point of view, even though scientific progress requires a debate of ideas.

Also read: How to do math problems with Gemini AI, easy and fast

Stuck in “Infinite Loop”

Lauren Williams, a mathematics professor from Harvard who was involved in this research, discovered another unique phenomenon. When tested with real research problems, AI often gets caught in a vicious cycle (infinite loop).

The AI will give an answer, then correct itself (“Wait, something’s wrong!”), give a new answer, correct it again, and so on without ever arriving at a final correct solution.

Tamara Kolda, one of the research authors from MathSci.ai, also warns of other, more fundamental dangers. According to him, AI actually has the potential to slow down scientific progress in the future.

The reason is, AI will only repeat the point of view that is told to it, in contrast to human colleagues who can debate and provide new, challenging perspectives, as compiled KompasTekno from The New York Times, Tuesday (10/2/2026).

Also read: 8 AI applications that can answer math questions, free and easy to use

KOMPAS.com is committed to providing clear, trustworthy and balanced facts. Support the sustainability of clear journalism and enjoy reading comfort without advertising through Membership. Join KOMPAS.com Plus now