When AI Solves Math Problems, What's Left for Human Mathematicians?

The world's greatest living mathematician just gave AI a qualified stamp of approval. Terence Tao of UCLA recently declared that AI has reached a point where it can "collaborate with human mathematicians." What earned this recognition? ChatGPT and other AI models have begun solving some of the 1,000+ unsolved mathematical questions known as the Erdős Problems, named after Hungarian mathematician Paul Erdős.

Over the past few months, researchers have made increasingly bold claims about AI's mathematical prowess. OpenAI President Greg Brockman celebrated on X in January: "GPT-5.2 Pro for solving another open Erdős problem. Going to be a wild year for mathematical and scientific advancement!"

The "Cheap Wins" Revolution

But when I spoke with Tao about these developments, his assessment was more nuanced than the headlines suggest. The AI-generated solutions are impressive, he told me, but represent what he calls "cheap wins." The bots are systematically working through the 1,000 problems and picking off the easiest dozen or so—a very different approach from how humans tackle mathematical challenges.

"Humans would not systematically go through all 1,000 problems and pick the 12 easiest ones to work on, which is kind of what the AIs are doing," Tao explained. The problems vary wildly in difficulty, and the ones AI has solved so far are those that an expert could have worked out "if they had half a day to look into the matter."

This reveals something fundamental about how AI approaches mathematics. While human mathematicians might spend years on a single profound problem, AI takes a more industrial approach—surveying the landscape and harvesting the low-hanging fruit with methodical efficiency.

The Helicopter Problem

Tao's most striking insight comes through a hiking metaphor. Traditional mathematical problem-solving is like taking a journey to a distant location, he explains. "You can lay down trail markers that other people could follow, and you could make maps." The process itself generates valuable insights that benefit the entire mathematical community.

AI tools, by contrast, are "like taking a helicopter to drop you off at the site. You miss all the benefits of the journey itself." This captures a crucial limitation: while AI can reach the destination, it doesn't provide the roadmap that makes mathematical progress truly valuable.

When human mathematicians tackle problems—whether they succeed or fail—they produce insights that others can build upon. AI-based proofs, however elegant, often lack this collaborative dimension that drives mathematical knowledge forward.

From Case Studies to Population Surveys

Despite these limitations, Tao sees transformative potential in AI's mathematical capabilities. He draws an analogy to medical research: "If you were to study a disease in the 18th century, you might study one patient and record all their symptoms. But in the 21st century, you can do a clinical trial with 1,000 people and get much more precise information."

Mathematics, he argues, is still operating at the case-study level. "A paper will take one or two problems and study them to death in a very handcrafted, intensive way." AI tools could enable "population studies" in mathematics—systematic exploration of problem spaces at unprecedented scale.

This shift could be particularly valuable for the tedious computational work that mathematicians typically try to avoid. "AIs will just happily blast through those tedious computations," Tao notes, potentially freeing human mathematicians to focus on more creative aspects of their work.

The Co-Author Timeline

Tao's predictions about AI's mathematical future have proven remarkably accurate. In 2023, he wrote that by 2026, AI would become a "trusted co-author" capable of contributing at the level of a technical paper co-author. That prediction drew mixed reactions—some called it too ambitious, others too pessimistic.

Today, he believes we're almost exactly on schedule. "We are basically seeing AIs used on par with the contribution that I would expect a junior human co-author to make, especially one who's very happy to do grunt work and work out a lot of tedious cases."

The Trust Problem

But significant challenges remain. The biggest issue, according to Tao, is AI's inability to accurately assess its own confidence. "When an AI gives you an answer, it does not give you any good indication of how confident it is, or it will always say, 'I'm completely certain that this is true.'" Human collaborators naturally express uncertainty, which provides crucial information for evaluating their contributions.

Tao also criticizes the AI industry's obsession with "push-of-a-button, completely autonomous workflows." For difficult problems, he argues, "you really want a conversation between humans and AI. And the AI companies are not really facilitating that."

The mathematical community faces a compressed timeline for establishing standards around AI use. "We've figured out computer-assisted proofs over 20 or 30 years," Tao explains. "Unfortunately, the timelines are much more compressed now. So we have to figure out our standards within a few years."