Science1 publisher3 min readPublished
Terence Tao names AI's arrival in math research a crisis in the field's values
Scientific American's account from the mathematicians' congress in Philadelphia has a Fields medalist quitting for OpenAI and a Millennium Prize problem falling to an opaque mix of models, with no verification trail reported for either.
The Scientist · Science desk
What happened
- Terence Tao used his July 24 lecture, "Mathematics in the Age of AI", as the centerpiece of the International Congress of Mathematicians in Philadelphia, the field's largest gathering.
- Days before he spoke, according to Scientific American, Anthropic's Claude Fable AI disproved the longstanding Jacobian conjecture with a tweet-length equation, shortly after ChatGPT cleared two other open problems.
- On the Tuesday of congress week, mathematicians solved the first Millennium Prize Problem in 20 years using what the magazine describes only as an opaque combination of Anthropic and OpenAI models.
- The medal winners' names had already leaked earlier that week, after someone asked ChatGPT to scrape the congress website.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint A result whose model pipeline is described as opaque cannot be refereed or attributed in the usual way, so the binding limit on absorbing these results is the field's checking capacity rather than any model's capability.
- decision Departments that priced a single hard theorem as a publication plus a tenure-track post now have to decide what that theorem is worth when its production cost has collapsed.
- exposure MIT's Andrew Sutherland puts every credential built on scarce, verifiable output in the same position as mathematics, calling mathematicians the canary in the coal mine for other professions.
- precedent Because mathematics is the proof point AI firms use with investors on superintelligence, the discipline should expect the pace of claimed results to keep being set by companies that owe it no verification standard.
Checking a disproof and checking a proof are not the same task. If a conjecture falls to one short equation, as Scientific American reports the Jacobian conjecture did to Anthropic's Claude Fable AI [3], anyone holding the definitions can substitute and check it; where the equation came from barely matters. A positive solution to a Millennium Prize Problem does not behave that way, and the account of that result names no artifact at all. The magazine says mathematicians reached it with "an opaque combination of Anthropic and OpenAI models" [5]. A pipeline that cannot be reconstructed cannot be refereed, and its credit cannot be apportioned either.
The career arithmetic in the same account is more legible than the mathematics. Scientific American reports that a few months earlier, any one of these feats would have secured a young mathematician a high-profile publication and a tenure-track job [6]. Fields Medals go to at most four mathematicians under 40, once every four years [7]. One of this cycle's winners, Jacob Tsimerman, used the winners' press conference to say he is leaving mathematics for AI safety work at OpenAI [8], which puts at least a quarter of the maximum cohort out of research mathematics on the day the medals were announced [9].
This account shows no institution changing a single rule. Journal policy, formalization in a proof assistant, a referee's verdict, and named human coauthors are all absent from the machine-assisted results [14]. Tao's contribution is a diagnosis and a question: he called the situation "a very confusing mess right now" and a crisis in mathematical values and practices [2], and told the hall that the field needs to step back and ask why it does mathematics in the first place [10]. That is a long way short of a renegotiated definition of authorship, and treating the two as the same thing would overstate what happened in Philadelphia.
The pressure driving this has a mechanism, and it sits outside the discipline. Scientific American reports that the AI industry believes conquering mathematics' objective truth is how it will convince investors that superintelligence is on the path, that several companies have made mathematics a top priority, and that one AI-for-math startup sells itself with "Solve math, solve everything" [11]. Mathematics is thus being run at a cadence set by firms that do not employ most mathematicians and do not owe them a verification standard.
That gap is where the real shortage is. On the evidence in this account, producing candidate results got cheap faster than checking them did, and the checking apparatus the field already has, referees and formal proof systems, is exactly what nobody in the story invoked. A discipline whose value proposition is certainty needs a logged pipeline and a machine-checkable proof object behind every claim. The three headline results described here arrived with neither.
What to watch
- Whether a machine-checkable proof object appears for the Jacobian counterexample or the Millennium result, and whose name is on it.
- Whether any journal or hiring committee publishes an actual rule on authorship and disclosure for model-assisted proofs.
- Whether other medalists follow Tsimerman into industry safety work, and what the next cycle's cohort looks like.