On September 11, 2026, Terence Tao published a blog post titled A severe misalignment on his personal site. The same day, The Economist ran a feature on AI and mathematics.
A Fields Medalist and a legacy publication, both pointing at the same thing on the same day: LLM performance on math doesn't match the hype.
The story blew up on Hacker News. The thread on mathandai.org hit 403 points with 467 comments, making the front page. For a topic this academic, that's unusual traffic.
I read through the blog post and the Economist piece and pulled out the parts worth talking about.
The Gap Between Competition Math and Actual Research
Here's the lay of the land.
LLMs are now scoring near-perfect or medal-worthy results on competitions like AIME and IMO. GPT-5, Claude Opus 4.5, and Gemini 3 Pro picking up silver at the 2025-2026 IMO isn't news; gold is what makes headlines.
But swap the problems for open-ended research questions, things like:
- Given a conjecture, construct a counterexample
- Given a 30-page proof, identify which step has suspicious "motivation"
- Given a mathematical field, decide which research direction deserves three years of effort
The models fall apart. Not "wrong answer" falling apart, but "sounds plausible and completely misses the point" falling apart.
Tao calls this class of problems "research-flavored but not research-grade." They look like research questions but are really competition problems wearing a research costume. Models can handle the latter, not the former.
What Tao Is Actually Criticizing
Tao isn't making a vague "AI is bad" argument. He targets specific weaknesses.
First, no sense of mathematical taste.
Mathematicians have strong judgment about which problems are beautiful, which directions have meat on the bone, which conjectures are worth pursuing. This judgment is implicit in papers, embedded in the author's decision to study this problem instead of that one. An LLM can read 1,000 papers and summarize the methods, but it can't explain why Grothendieck chose schemes over the alternatives.
Second, no ability to judge proof elegance.
The same theorem can have five proofs. One might be a brute-force slog spanning 50 pages; another might wrap up elegantly in two. Tao's point: LLMs can tell you whether a proof is correct, but not whether it's good.
Third, no ability to judge whether a research path is worth pursuing.
This is the killer. The core skill in research is knowing what to abandon. Ninety percent of attempted directions are dead ends; mathematicians filter them out through experience, intuition, and taste. LLMs don't have an "abandon" function. They generate a plausible-looking plan for every path.
Tao argues these three gaps explain why "model wins gold at IMO" and "model functions as a research assistant" sit on opposite sides of a nearly impassable canyon.
Why This Hurts AI for Science Even More
Mathematics is the cleanest testbed in AI for Science.
Why? Because:
- The data has no noise, unlike biology experiments with batch effects or materials synthesis with equipment variability
- The objectives are verifiable: a proof is correct or it isn't, no "close enough"
- The feedback signal is clear: the Lean compiler checks in seconds
In other words, if AI can't handle open-ended problems in the cleanest domain of all, it will only do worse in biology, materials science, and earth science, where the data is messy, the goals are fuzzy, and feedback takes forever.
Over the past two years, "AI for Science" has been one of the sexiest slide-deck topics in AI. Post-AlphaFold, everyone pitches "AI accelerating research." But what Tao highlights is this: benchmark scores ≠ research value. A model that scores a perfect AIME and an assistant that helps you push your PhD project forward are different species entirely.
This point got reinforced again.
The Overlooked Upside: Where AI Actually Helps Mathematicians
Tao doesn't deny AI is useful. On the contrary, he mentions in his blog that he regularly uses AI for:
- Lean code assistance: formal verification is painful, and AI can produce a first draft
- Literature search: with 200+ new math papers hitting arXiv daily, AI handles summaries and classification well
- Grammar checking, LaTeX debugging: tedious tasks where AI beats humans
The misalignment is this: current products and benchmarks position AI as a "mathematician," but it's actually a "faster calculator plus search engine."
That's the real problem: mispositioning.
Concrete example: if you use Lean plus AI for formal verification, you'll find LLMs are genuinely helpful. Formal languages have strict syntax, compiler feedback, and clear objectives. That's their sweet spot. But if you ask the same model to "understand the motivation" in an Erdős paper, you'll likely get a grammatically perfect summary that misses the entire point.
AI isn't a mathematician. AI is a mathematician's IDE. The market deliberately blurs this positioning because nobody writes a check for "IDE," whereas "AI mathematician" gets VCs excited.
Editorial Take
Why does this matter? Not because "Tao said AI is bad," but because he used a specific, verifiable domain (mathematics) to puncture a widely accepted narrative (AI = research assistant).
My read: this discussion will ripple well beyond mathematics. Three reasons.
First, math is the benchmark case. If AI can't handle "taste" and "abandonment" in the cleanest domain, the whole AI for Science narrative needs rewriting. This directly affects how NIH, DARPA, and other national funding bodies allocate grants; they're actively weighing "AI-accelerated research" directions.
Second, evaluation frameworks will get blowback. The gap between benchmarks with standard answers (AIME, IMO) and actual research capability will become obvious to people in other fields. Coding benchmarks like SWE-bench and HumanEval measure the same thing as AIME: can you do the task, not whether you should. Products like Devin will face the same scrutiny.
Third, product positioning will get forcibly reset. Tao's "AI is a faster calculator plus search engine" line will get quoted repeatedly. Over the next 18 months, every "AI Scientist" or "AI Researcher" product will have to answer the same question in their marketing: can your model decide which path to abandon?
How is this different from past events? Over the past two or three years, "AI surpasses humans in X" headlines have landed every few weeks (AlphaFold, IMO gold, ICPC champion). But Tao's critique comes from inside the field, from a working top-tier mathematician, not from a competitor AI company. That kind of criticism carries different weight than, say, an Anthropic CEO saying "we need to be careful."
What remains uncertain: Tao didn't propose an operational benchmark for "how to evaluate AI research capability." He said "research ability ≠ competition scores" but didn't answer "then what does it equal?" That gap will get filled by various players, and what they fill it with will determine where the next round of AI for Science money flows.
Sources:
- Tao's blog post A severe misalignment: terrytao.wordpress.com/2026/09/11
- The Economist feature on AI and math: economist.com/science-and-technology/2026/09/11
- Hacker News thread: item?id=49662371
- Event aggregation page: mathandai.org