Terence Tao published a blog post titled "A digestion of the Jacobian conjecture counterexample" on his WordPress blog on July 21, 2026 (Tao's blog). The post walks through a counterexample to a long-standing problem in algebraic geometry and commutative algebra, and it documents Tao's use of ChatGPT during the process.

The Jacobian conjecture, proposed by Ott-Heinrich Keller in 1939, concerns polynomial maps — functions built from sums and products of variables, like 3x²y + y³. The conjecture says that if such a map, from n-dimensional complex space to n-dimensional complex space, has a constant nonzero Jacobian determinant (a particular measure of how the map distorts space at each point), then the map must be invertible, and the inverse must also be a polynomial map. The conjecture has resisted proof for decades and is notorious for the number of claimed proofs and counterexamples that have failed under scrutiny. A confirmed counterexample would settle the problem definitively in the negative.

Tao's post does not claim to verify or refute the counterexample. The title frames the work as a "digestion" of the result, suggesting an effort to internalize and communicate the mathematical content rather than to serve as a formal referee report. His blog has long functioned as a venue where Tao works through mathematics in a semi-public, exploratory register, and this post fits that pattern.

What sets this post apart from Tao's usual exposition is the explicit role of ChatGPT in the mathematical work. A commenter on the post observed that ChatGPT did not need to be corrected even once during the conversation (Tao's blog). That observation, appearing in the comment section of a post by a Fields Medalist working on a problem of this caliber, is worth pausing on.

The verified facts support a narrow claim: Tao used ChatGPT as a conversational partner while digesting the mathematics, and the interaction proceeded without Tao needing to correct the model's output. What exactly ChatGPT contributed — whether checking algebraic manipulations, reconstructing arguments, or serving as a sounding board for exposition — is not specified. The absence of corrections, noted by a third-party commenter rather than by Tao himself, is a single data point, not a systematic evaluation.

The context matters here. The Jacobian conjecture sits in commutative algebra and algebraic geometry, domains that demand precise symbolic reasoning, careful handling of ring-theoretic structure, and fluency with polynomial ideals and their behavior under transformations. These are not tasks where language models have historically excelled. Early-generation models routinely made algebraic errors that a competent graduate student would catch immediately. The fact that a mathematician of Tao's caliber engaged ChatGPT on this material and encountered no errors worth correcting in-line is a signal about where these tools now sit on the curve of mathematical utility.

It is also a single interaction, reported by a commenter, on a specific mathematical task. Generalizing from it would be premature. Mathematicians have been experimenting with LLMs for proof assistance, literature search, and informal exploration for the past couple of years, and the results have been uneven — useful in some contexts, unreliable in others, with the reliability boundary itself poorly understood. Tao has been notably open about experimenting with AI tools in his research workflow, and this post is consistent with that pattern of public, pragmatic exploration rather than advocacy.

The broader question the post raises, without explicitly asking it, is whether LLMs are crossing a threshold in mathematical conversation. Not yet replacing formal proof assistants like Lean or Coq, which operate at the level of machine-checked certitude, but becoming reliable enough for a working mathematician to treat as a genuine interlocutor on hard problems. The distinction between "no corrections needed" and "output is correct" is doing a lot of work in that sentence, and the available facts do not resolve it. What we have is an observation, not an evaluation.

In my view, the most interesting aspect of this story is the ordinariness of the interaction. Tao used a language model while working through mathematics, noted it in a blog post, and a commenter remarked on the absence of errors. No press release, no benchmark, no demo. The tool is becoming part of the workflow, and the workflow is becoming part of the documentation. That is a quieter and more durable signal of adoption than any capability announcement.

The counterexample to the Jacobian conjecture, if it holds, resolves one of the field's persistent open questions. Whether ChatGPT played a substantive role in understanding it or was merely a clean conversational mirror remains an open question of its own.