Every maths teacher says the same thing, and every child finds it infuriating.

Show your working.

Not because the answer doesn't matter, but because an answer you can't check isn't yet a result. It's an assertion.

On 21 September 2026, OpenAI published a short post about forming an advisory group of mathematicians. Its opening paragraph contains this:

The Millennium Prize problem is in a subordinate clause. In addition to resolving it. It is framed as the smaller part.

The Navier–Stokes existence and smoothness problem is one of the seven Clay Millennium Prize Problems, set in 2000, each carrying a $1 million prize. Six remain unsolved. The equations govern how fluids move - water through a pipe, air over a wing, blood through an artery, weather across a continent. We use them constantly and cannot prove that their solutions always behave.

Twenty-five years of the world's best mathematicians have not settled it.

OpenAI says a model it started training on 28 August did, along with a hundred other open problems. The announcement came 24 days later.

They also write that the pace "has surprised the mathematicians within OpenAI" - an unusual thing for a company to publish about its own product, and either disarming honesty or very good copywriting.

No proof. No paper. No preprint. No list of which hundred problems. No named reviewer. No external verification of any kind. The model is internal and unreleased, so no mathematician outside OpenAI can probe it.

The honest word for all of this is unverifiable.

Not false. We want to be careful here, because "unverified" gets read as "debunked" and it isn't. A hundred resolved problems is not a thing a company fabricates wholesale, and OpenAI has real mathematicians and a real track record in this area. This may turn out to be exactly what it says.

But right now, the entire claim rests on OpenAI's word about a model nobody else can use, concerning proofs nobody else has read.

In mathematics of all fields, that is not how anything becomes true. Mathematics is the one discipline where verification is total: a proof is either valid or it isn't, and anyone competent can check. That is the whole point of it. A mathematical claim withheld from checking is the most incomplete kind of claim there is.

Here is the part that raises this above a press release, and OpenAI gets credit for including it.

The announcement links to an open letter called "A Severe Misalignment of AI in Mathematics," in which mathematicians raise concerns about "the negative externalities of solving open problems as a benchmark for new AI systems."

The complaint underneath that phrase is worth unpacking. Open problems are not a scoreboard. They are the accumulated attention of a discipline - the places where mathematicians have decided the interesting difficulty lives. Treating them as benchmarks to be cleared by a model changes what they're for, and risks converting a field's intellectual agenda into a vendor's evaluation set.

A company linking to the argument against itself is doing something most don't.

OpenAI's response is an independent advisory group - and on its own telling, OpenAI did not create it. It is "working with mathematicians who have established an independent mathematics advisory group," hosted at the Institute for Advanced Study.

The terms are stronger than industry advisory boards usually get:

In OpenAI's words: "Its value depends on its members being able to exercise their own judgement and challenge ours."

The membership matches the terms. François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, Melanie Matchett Wood. Gowers, Hairer and Witten hold Fields Medals - that's our note, not OpenAI's. These are not people who attach their names to window dressing, and the IAS is not an institution that hosts it.

So take the independence seriously. Then read the last line of that paragraph:

They can assess significance. Advise on dissemination. Comment on professional standards. Criticise OpenAI in public, unprompted, in writing.

They will not advise on speed.

The scope of an oversight body is where its real authority gets decided, and it is normally settled quietly, in terms nobody publishes. Here it is stated outright, which is to OpenAI's credit even as it tells you exactly where the boundary is.

And there's a defensible case for it. Pace is a business decision, not a mathematical one. A group convened to advise on mathematical significance has no particular standing to run a company's research schedule.

But notice what pace governs. Whether results arrive faster than they can be reviewed. Whether a field gets time to adapt - which is precisely the concern in the open letter OpenAI linked. Whether "we were surprised by how fast it went" is a sentence anyone is positioned to act on.

Every other question the group can address is downstream of the one it cannot.

This is now the third story we've covered in a month that turns on the same axis. Anthropic's CEO published an essay urging the industry to pace itself. Four customers sued four AI companies for allegedly agreeing to do so. And now an oversight group with real teeth has been constituted with speed explicitly outside its remit.

Nobody has worked out who is allowed to say "slower."

Nothing here is about children, and the connection below is ours.

Your kid is going to see a headline saying AI solved a million-dollar maths problem. It will be presented as settled. It isn't yet - and the gap between "a company announced it" and "the field has checked it" is the single most useful thing you can teach about how knowledge actually works.

Show your working is not a school rule. It's the rule. It applies to a seven-year-old's subtraction, a scientist's paper, and a hundred-billion-dollar company's press release, and it applies most strictly when the claim is most impressive.

Some ways to use this at the dinner table:

If a meaningful fraction of these results survive review, this is one of the most significant scientific events of our lifetimes, and the caution in this piece will read as fussy.

We'd be glad to be fussy. The claim is still the kind that has to be checked, and it has not been.

Until then: show your working.

Source: OpenAI, "Advisory Group on Mathematics and Artificial Intelligence," 21 September 2026. https://openai.com/index/advisory-group-on-mathematics-and-ai/

This is a company announcement, not independent reporting. No proof, paper, preprint, problem list or external verification accompanied it, and the model referenced is internal and unreleased. Quotations are verbatim. The Fields Medal attributions, the analysis of the pacing carve-out, and the guidance for parents are ours.

Disclosure: OpenAI is a direct competitor of Anthropic, and this package was drafted using Claude, made by Anthropic.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter