Terence Tao sees a future in which AI changes how mathematical research gets done. In an interview with The Atlantic, the Professor of Mathematics at UCLA described a shift from solitary, long-running work toward larger collaborations supported by chatbots, proof assistants, and other computer tools.
The idea is not that AI replaces mathematicians. Tao’s view is more specific: AI could become part of a research workflow that lets people explore more directions, test more ideas, and organize broader projects than an individual expert could manage alone.
From Individual Depth to Team Scale
Tao calls this possibility “industrial-scale mathematics.” The phrase points to a different way of organizing research. Instead of one expert spending years on a narrow set of difficult problems, larger groups could use AI support to pursue many related paths at once.
That kind of work would be broader and, in Tao’s description, less deep than the traditional model. But it could still produce useful mathematical insight. The value would come from scale: more people, more assisted exploration, and more ways to convert rough ideas into forms that computer tools can check.
Chatbots are central to this vision because they can act as a bridge between informal human thinking and formal systems. Tao imagines researchers talking with AI assistants to develop and refine ideas. Those assistants could also help translate natural language into code for proof assistants.
That translation step matters because a mathematical idea often has to be made precise before it can be tested in a formal system. If AI can reduce the friction between a mathematician’s explanation and the code needed by a proof assistant, it could make larger coordinated projects easier to run.
Why Tao Compares AI to Chess Computers
Tao’s comparison is chess. Chess computers mastered the game, but chess did not disappear. Instead, players now use engines to analyze moves far ahead and study positions in ways that were not available before.
For mathematics, the analogy suggests that AI could become a powerful tool without ending the human side of the field. A chess engine can calculate deeply, but the game still has players. In the same way, AI systems may help mathematicians examine possibilities, check work, or handle routine tasks while humans continue to set direction.
OpenAI’s o1 shows promise in this direction, according to Tao. But he does not describe current AI models as true research assistants. His assessment is more cautious: they are “mediocre, but not completely incompetent” assistants that can perform routine work but still lack creativity and flexibility.
That distinction is important. A useful assistant can save time, but a research collaborator must do more than complete isolated tasks. Mathematical research often depends on noticing when an approach is wrong, changing strategy, and building judgment over time.
The Learning Gap Between AI and Researchers
Tao draws a sharp line between AI systems and graduate students. “One key difference between graduate students and AI is that graduate students learn,” he explains.
In his view, an AI model may adjust its behavior for a while after being corrected. But it often returns to earlier methods, and each new AI session begins again from scratch. That makes failure less productive than it is for a human learner.
Tao says he is “much more patient with graduate students because I know that even if a graduate student completely fails to solve a task, they have potential to learn and self-correct.” He also stresses that “AI and humans have such different models for how they learn and solve problems.”
This is one reason the future he describes is collaborative rather than fully automated. AI can process enormous amounts of information, but the human role remains central when research requires judgment, intuition, and a willingness to rethink the problem itself.
“AI is very good at converting billions of pieces of data into one good answer. Humans are good at taking 10 observations and making really inspired guesses,” Tao notes.
A Proof Example Shows Both Promise and Limits
The source also points to a concrete example involving Robert Ghrist, a mathematics professor at the University of Pennsylvania. Ghrist used GPT-o1-mini to create a complex mathematical proof after months of testing various models.
In that case, GPT-o1-mini analyzed a faulty proof, found errors, and produced a new, more elegant proof. That is exactly the kind of workflow that makes AI interesting for mathematical research: the model was not simply generating text, but working through an existing mathematical attempt and improving it.
Still, the result also showed the limits of the process. Ghrist said that working with the AI did not necessarily make the work simpler. Later, another mathematician showed that the proof could have been much simpler.
That outcome fits Tao’s broader point. AI may help expand what researchers can try, and it may assist with formal or routine parts of the work. But the simplest, most insightful route may still come from human mathematical judgment.
What Industrial-Scale Mathematics Would Mean
If Tao’s vision develops, AI mathematics would not just be about faster answers. It would be about changing the structure of research itself.
The practical shift would involve several connected roles:
- Researchers using chatbots to discuss and refine ideas.
- AI systems helping convert natural language into code for proof assistants.
- Large teams exploring broad, crowdsourced problems with AI support.
- Humans continuing to supply creativity, direction, and inspired guesses.
That combination is why Tao describes the approach as complementary. Narrow, deep mathematics would still matter. But alongside it, AI could support broader projects that gather insight at a much larger scale.
The current systems are not there yet. Tao’s comments make clear that today’s models remain limited as research assistants. But the direction is significant: mathematics may become one of the fields where AI is most useful not as an oracle, but as infrastructure for a different kind of collaboration.