Can AI Revolutionize Mathematics? A Science Friday Conversation on AI, Proofs, and the Future of Math

Science Friday host Flora Lichtman talks with mathematicians Emily Real and Daniel Litt about AI's impact on mathematics. They trace AI's progress from basic problem solving to undergraduate proofs and now toward research-level questions, while debating whether AI could ever fully revolutionize the field. The guests emphasize that AI is changing how mathematicians work, but there is no single AI breakthrough that redefines the discipline. They discuss the potential of proof assistants to formalize proofs, the risk of AI-generated nonsense in preprints, and how AI might augment, rather than replace, human creativity and collaboration. The conversation offers both cautious optimism and a sober cautionary note about trust, understanding, and the human elements of mathematical discovery.

Overview

In this episode, Science Friday hosts a thoughtful dialogue with two seasoned mathematicians, Dr. Emily Real from Johns Hopkins University and Dr. Daniel Litt from the University of Toronto, about whether artificial intelligence is truly transforming the practice of mathematics. The guests begin by charting AI's trajectory through the learning curve of mathematics: from solving problems with numerical answers in school to tackling contest-level proofs and, more recently, being tested on research-level questions. They acknowledge the impressive progress of AI in math yet push back on the idea that machines have already rewritten what it means to do mathematics. The discussion situates AI as a powerful tool that is reshaping workflows, collaboration, and the pace of discovery, while leaving room for human insight, taste, and the long arc of mathematical imagination.

Quote 1 "the specific claim that AI tools are revolutionizing mathematics definitely has not come true yet" - Daniel Litt, University of Toronto

AI in School and Research Problems

The conversation traces AI's progress from classroom problems with definite numerical answers to undergraduate contest-style proofs, and then to open, research-centric questions that intrigue paid mathematicians. Emily Real stresses that AI’s current strength lies in solving problems that historically required human computation, and improvements over the past year have made AI capable of handling more complex tasks than before. Daniel Litt concurs that there is evidence of meaningful, even significant, progress, but he cautions that the leap to a full-scale revolution in mathematics still hasn’t happened. The speakers agree that we should calibrate expectations and recognize the incremental, cumulative nature of AI’s influence on the field.

Quote 2 "leading indicators that it will be very significant" - Daniel Litt, University of Toronto

Formal Proofs, Verification, and Proof Assistants

The discussion then pivots to the idea that mathematics isn’t only about proving theorems but also about stating problems, identifying interesting questions, and developing a theory around them. A core theme is how AI might integrate with formal proof systems. Emily explains that a major advantage of AI in math could be its ability to generate and organize proofs within the language of computer proof assistants, which can check logical steps rigorously. This formalization could help automate parts of the refereeing process, reducing the burden on humans and increasing reliability. The guests emphasize that trust in a proof is not fully earned by natural language explanation alone; it should be accompanied by formal, machine-checked steps when necessary.

Quote 3 "trusted software programs called computer proof assistants that are engineered by human experts, so not produced with AI in any way that will take a very precisely written mathematical proof and check the logic line by line" - Flora Lichtman

Proofs, Oracles, and the Two Modes of AI Proving

The hosts discuss the provocative idea that AI could someday function as an oracle, offering correct statements that humans would then seek to understand and re-prove in more transparent, human-accessible ways. Daniel acknowledges this potential but notes that the value lies not just in correctness but in the ability to explain. Flora adds that there are two modes by which an AI could produce a proof: one that presents a long, fully formal, machine-checkable argument, and another that delivers a natural-language narrative that still requires human formalization and verification. The risk of accepting AI outputs without sufficient justification is highlighted, particularly given the trust issues surrounding language models that do not possess beliefs or theory of mind.

Quote 4 "there are sort of two different modes in which an AI can produce a proof" - Flora Lichtman

Vibe Proving, Hallucinations, and the Need for Human Judgment

The panel turns to practical workflow questions, notably the phenomenon of AI ‘hallucinations’ in mathematics. Flora recounts the real-world risk of AI-generated papers containing nonsense, citing an example where many Hodge conjecture papers on a major preprint server were not credible. She compares AI-driven speculative outputs to human missteps in early stages of problem solving and emphasizes that verification and validation remain essential. The participants discuss how AI could enhance the mathematical workflow by suggesting directions or providing partial proofs, while still requiring a careful human reviewer to determine the validity of results.

Quote 5 "11 of them were nonsense generated by AI tools like hallucinations" - Flora Lichtman

Creativity, Collaboration, and the Future Role of AI in Math

Beyond technical capability, the speakers discuss what makes a great mathematician. They argue that mathematical progress often arises from collaboration, dialogue, and the exchange of ideas at chalkboards, rather than solitary problem solving alone. Flora remarks that the most influential mathematics blends technical skill with philosophy, aesthetics, and a sense of structure, and she asserts that time and dedication are crucial to mastery. She also reflects on how AI could expand the boundaries of what a mathematician can learn by assisting with literature review, staying abreast of developments, and enabling deeper exploration of new fields such as hyperkähler geometry. The dialogue emphasizes that AI is likely to shift the skill set in mathematics—enhancing some capabilities while making others less central—without erasing the value of human insight and curiosity.

Quote 6 "creativity often arises in conversation between mathematicians" - Flora Lichtman

Closing Reflections

The episode closes with a shared sense that AI tools will become integrated parts of mathematical practice, accelerating workflows and opening new possibilities, while also introducing risks around verification, trust, and interpretability. The guests agree that we are in an era of cautious optimism: AI will not instantly redefine mathematics, but it is already reshaping how mathematicians think, collaborate, and approach unsolved problems. The takeaway is that the best path forward combines human mathematical intuition with AI-powered tools, guided by careful verification and a commitment to clear, formal explanations when needed.