Can AI Really Do Math, or Is It Just a Very Fast Calculator?
AI·November 15, 2026
A new test is putting one of the most argued-over questions in AI to work: can today's models do real mathematics, or do they mostly crunch numbers very quickly? In a recent episode of Scientific American's podcast Science Quickly, host Kendra Pierre-Louis and SciAm reporter Joe Howlett looked at a challenge designed to probe that question with harder material than typical benchmarks. The aim is to see whether AI systems can tackle cutting-edge proofs, the open-ended reasoning that working mathematicians do, rather than problems with answers already known.
The episode walks through how the experiment is set up and what early results suggest. As the discussion frames it, those results are mixed, which is part of why they are hard to read as a clean verdict in either direction.
The bigger dispute is over what any progress actually means. One camp sees AI as potentially changing how mathematics gets done, with new kinds of discovery becoming possible. A more cautious camp argues that these tools are powerful assistants that speed up existing work without yet originating ideas on their own. The episode presents the question as still open rather than settled.
For readers outside mathematics, the stakes are practical. Proofs are a demanding test of reasoning, so how AI performs on them offers a useful signal about how much weight these systems can bear in other complex analytical work. Whether the answer turns out to be "revolution" or "high-powered assist" will likely depend on how the next rounds of testing play out.