Toby Ord: AI Automates Proof, but Math Is Much More

On August 23, Oxford researcher Toby Ord published "Mathematics is Much More than Proof," responding to claims that AI will replace mathematicians. He concedes that automated proof is clearly underway but argues that math is far more than proof, and that the ability to ask the right questions remains unproven as an AI capability. David Manheim joined the debate the next day, producing an initial exchange over LLMs' question-asking abilities.

Confirmed

• Ord noted that symbolic computation was automated over the past forty years, and we may now be witnessing the automation of mathematical proof (or most of it), analogous to how 20th-century computers automated calculation. Per m9, he noted AI in 2026 can already prove results ranging from conjectures to major theorems (e.g., OpenAI-related cases).
• He stressed that much of mathematical work, as in science, lies in figuring out what questions to ask. From an information-theoretic view, a question carries nearly as much information as the proposition it explores (a yes/no question is one bit less).
• He illustrated with an example: even a machine that could prove any true mathematical statement in microseconds would be useless if it could not conceive of problems like the unit-distance problem before all stars burn out.

Unconfirmed

• Ord stated there is little evidence either way on whether AI can replicate humans' unique ability to find important questions among exponentially many candidates; he also admitted human question selection is uneven and biased.
• He identified the key bottleneck as the lack of good training data (or verifiable rewards) for "asking good questions"; unless the skill comes for free from pretraining on mathematical texts or learning proofs, it is hard to see why models would have it.
• David Manheim agreed that good questions carry substantive information but, citing mathematicians' vague impressions, said LLMs already perform quite well at identifying interesting research directions and proposing hypotheses—tension with Ord's caution, though neither side offered firm evidence.

Why it matters

• The exchange offers a finer-grained view of the "will AI replace mathematicians" debate: automated proof is clearly coming, but higher-level work like research taste and question selection may either remain a durable human domain or become a key frontier for next-generation AI research.

2026-08-23 ~ 2026-08-24 · 9 related posts

Full story(3 episodes)→

• Episode 1: Mathematicians Debate "End of Mathematics": AI Superhuman Proofs Still Far From Humanless Math(2026-08-23, 5 posts)
• Episode 2: Toby Ord: AI Automates Proof, but Math Is Much More(2026-08-23, 9 posts)
• Episode 3: Debate: Should GPUs Take Over Hard Math Proofs?(2026-08-24, 2 posts)

• Mathematics is About Much More Than Proof: Reflections on AI Automation — tobyordoxford · 2026-08-23
• [source] Automating Mathematical Proof: AI Advances and the Future of Math — tobyordoxford · 2026-08-23
• The Importance of Asking the Right Questions in Mathematics — tobyordoxford · 2026-08-23
• Core math research lies in determining which questions to ask — tobyordoxford · 2026-08-23
• AI not yet proven capable of posing key questions like human mathematicians — tobyordoxford · 2026-08-23
• Lack of training data limits AI's ability to ask good questions — tobyordoxford · 2026-08-23
• Toby Ord: AI Automating Proofs Does Not Replace Mathematicians' Core Work — tobyordoxford · 2026-08-23
• [source] Debate on whether current LLMs can formulate math conjectures better than humans — tobyordoxford · 2026-08-24
• [source] Experts discuss LLMs' ability to identify math research directions — davidmanheim · 2026-08-24