Tao’s warning is less about machines replacing mathematicians than about confusing a verifiable answer with a worthwhile contribution. As proof generation gets easier, the field will need sharper ways to judge what a result adds.
Summary
Terence Tao’s Math 2.0 slides argue that mathematics may be moving from proof scarcity toward proof abundance, while stressing that most frontier problems remain unsolved. The practical tension is that proof assistants such as Lean can check whether a formal argument follows from its assumptions, but cannot establish that the question or result matters. Tao suggests that incentives centered on solved problems may become misaligned if producing solutions gets cheaper.
What happened
Tao describes proofs as unusually amenable to AI: they can be represented digitally, assessed against shared standards and, in suitable languages, mechanically checked. The slides point to Lean as an example, not as a claim that all mathematics can now be formalized or automated. Tao’s concern is what happens if AI systems produce more candidate solutions: a count of completed problems may become a weaker proxy for broader research goals. He calls for transparent reporting of resources and negative results, and identifies First Proof as one effort in that direction. The presentation does not show that research institutions have already changed their incentives.
What the archive adds
The archive’s prior coverage of Math 2.0 likewise distinguishes possible proof abundance from an exhausted mathematical frontier. According to the dossier, DeepMind reports converting about one million natural-language problems into roughly 80 million Lean instances for work on olympiad-level reasoning. This illustrates how formalized data and Lean’s checkable feedback can support search and reinforcement learning. Separately, DeepMind reports an IMO 2025 gold-medal-standard result using Gemini Deep Think; that competition result is not the same evaluation setting as formal proof generation. Neither establishes that research incentives have changed.
Trend signal
AI mathematics is developing along complementary tracks: formal systems that can verify proofs, and general reasoning systems tested on competition problems. The former can make correctness more auditable; the latter can demonstrate problem-solving without, by itself, showing that a result is formally verified or useful to research. The broader trend is a widening gap between what systems can produce and what institutions can reliably evaluate. Tao’s point is that machine-checkable correctness settles one question, not the question of mathematical value.
Why now
According to the dossier, one driver is the combination of search-guided reinforcement learning with Lean’s machine-checkable feedback; DeepMind reports a large pipeline of natural-language problems converted into Lean instances. A second is the expansion of formalized mathematical data and infrastructure, which gives systems more material on which to train and test. A separate driver is investment in general reasoning approaches: DeepMind reports an IMO gold-medal-standard result using parallel exploration. Tao’s primary slides add that proofs are digitally representable and mechanically checkable in suitable languages. These advances explain why the question is timely, but do not prove a broad shift in research incentives.
Connections to other events
The second-order effect may be a change in what researchers must evaluate. If verified solutions become easier to generate, the scarce judgment could be whether a result opens a new path, connects ideas or answers a worthwhile question. That is an inference, not a demonstrated change in mathematical practice. Still, Tao’s call to report resources and failures matters because it can help readers distinguish a reproducible contribution from a polished, selectively presented success. Formal checking protects against some errors; it does not supply mathematical taste.
Why it matters
Builders should treat proof checking as one layer of evaluation, not the whole scorecard. Preserve the formal statement and checker result, but also document the method, resource use, failed attempts and why the problem matters. Buyers and research leads should resist ranking systems only by the number of correct answers: that rewards yield while leaving significance and understanding unmeasured. A useful assessment should separate correctness, reproducibility and contribution rather than collapsing them into a single success count.
What happens next
- A public AI-math assessment will include a proof-assistant verification result for at least one evaluated formal problem set. (6 months, 70%). Watch for: A paper or benchmark release reports kernel-checked proofs for a named set of problems, rather than only judging final answers.
- At least one public AI-math evaluation will report resource use alongside its results, improving the basis for comparing demonstrations. (6 months, 55%). Watch for: A research paper or benchmark report publishes compute or tool-use details for its evaluated runs.
Limits and uncertainty
The primary evidence is Tao’s presentation, not an independent audit of current AI capabilities. A Lean-checked proof is correct relative to its formal statement and foundations; formalization itself can omit context or encode an uninteresting question. Competition performance and formal proof results are different kinds of evidence and should not be treated as interchangeable. The claim that solution-focused incentives may misalign is a plausible warning, not proof that journals, grants or hiring already reward the wrong things.
Skill to build now
Practice separating correctness from significance. This week, choose one short AI-generated proof claim: write its formal statement, check it in Lean if feasible, then add a two-sentence note explaining what the result contributes and what the checker cannot tell you.
Reader takeaway
This week, take one AI-generated mathematical claim and separate three checks: whether its statement captures the intended problem, whether a proof assistant verifies the formal argument, and what the result contributes beyond solving that statement. Write down a reason for its mathematical value that does not simply repeat that it is correct.
Open question
If a proof is machine-checked but its significance is disputed, who should decide whether it counts as a major mathematical contribution—and how should that judgment affect credit?