LLMs can write proofs that sound right. The Riemann case shows why that's dangerous

5 min read 1 source clear_take
├── "LLMs produce the shape of mathematical proofs without the substance — fluent prose is not proof"
│  └── Fernando Borretti (borretti.me) → read

Borretti works through the claimed Riemann hypothesis proof line-by-line and shows it is a sequence of plausible-sounding mathematical sentences arranged in the shape of a proof, not an actual proof. Steps are asserted rather than derived, cited lemmas don't exist or don't say what's claimed, and the crucial moves are always one level of abstraction away from being checkable.

├── "The real danger is epistemic asymmetry — generating fake proofs is cheap, verifying them is expensive"
│  └── top10.dev editorial (top10.dev) → read below

The editorial argues the surface story about Riemann masks a deeper problem: producing a 2,000-word document that reads like a mathematical paper now costs a few API dollars, while careful verification requires specialist time. This asymmetry means non-specialists can't distinguish sophisticated garbage from real work, and the same cycle has already played out with claimed P vs NP and Goldbach results.

└── "The HN audience split reveals the core problem — sophisticated nonsense fools non-specialists"
  └── @_dain_ (Hacker News, 183 pts) → view

The 183-point HN thread submitter surfaced Borretti's rebuttal into a discussion that was itself split between readers who immediately recognized the proof as nonsense and readers who were genuinely uncertain. That split is the real signal: the LLM output invoked zeta functions and analytic continuation in locally coherent paragraphs convincingly enough that even technical readers couldn't all tell it apart from real mathematics.

What happened

Fernando Borretti's 'Re: Riemann' is a line-by-line response to a widely circulated claim that a large language model had produced meaningful progress on the Riemann hypothesis — one of the six remaining Millennium Prize Problems and arguably the most famous open question in mathematics. The original post framed the model's output as a candidate proof worth taking seriously. Borretti, with the patience of someone who has seen this movie before, works through the argument and shows it is not a proof at all. It is a sequence of plausible-sounding mathematical sentences arranged in the *shape* of a proof.

The HN thread (183 points at time of writing) is split between people who already knew the 'proof' was nonsense and people who were genuinely uncertain — which is itself the interesting part. The problem isn't that the LLM wrote garbage. The problem is that it wrote garbage sophisticated enough that non-specialists couldn't tell. The output invokes zeta functions, functional equations, and analytic continuation in grammatically correct, locally coherent paragraphs. The connective tissue — the part where you actually have to prove something — is where it falls apart. Steps are asserted rather than derived. Lemmas are cited that don't exist or don't say what they're claimed to say. The crucial move is always one level of abstraction away from being checkable.

Borretti's piece is not the first of its kind, and it will not be the last. A near-identical cycle played out earlier this year with a claimed P vs NP result, and before that with Goldbach. The pattern is now well-established enough to deserve a name.

Why it matters

The surface story is about the Riemann hypothesis. The real story is about epistemic asymmetry under abundant generation. Producing a 2,000-word document that reads like a mathematical paper now costs essentially nothing — a few API dollars and some prompting. Reading that same document carefully enough to confirm it is wrong requires a professional mathematician's time, which is expensive and in short supply. The ratio has flipped. For most of history, writing a plausible proof was harder than reading one. That is no longer true.

What LLMs have made free is not understanding. It is the fluent *performance* of understanding. And fluent performance is, until very recently, how human expertise was signaled. We trusted writing that sounded confident and used the right vocabulary because producing that kind of writing used to require having read the literature. The correlation between register and competence has been severed, and nobody has rebuilt the replacement yet.

The HN thread surfaces a second, subtler observation that keeps coming up in these debates: the outputs have a stylistic fingerprint. Not a specific tic — a *gravitational pull* toward certain phrasings, certain hedges, certain ways of gesturing at technical machinery without engaging it. One commenter noted that Claude, GPT, and Gemini all produce suspiciously similar mathematical prose when pushed outside their competence. The homogenization is itself the tell. Real mathematicians have idiosyncratic styles; the models have converged on a kind of average-professor voice that reads smooth and says little.

The response from the AI-optimist camp is usually some version of 'but the models are getting better, and this will be fixed.' That may well be true for routine problems. It is almost certainly not true for the Riemann hypothesis in the next model generation, because the gap between 'can produce text about X' and 'can produce novel mathematics about X' is not a scaling gap. It is a capability gap that current architectures do not obviously close. Treating the failure mode as a bug to be patched, rather than a structural feature of the technology, is itself a category error.

What this means for your stack

The Riemann story is a mathematical curiosity, but the pattern it exposes is operational and it is already in your repo. Substitute 'proof of the Riemann hypothesis' with 'pull request that refactors the auth layer' and the shape is identical: a confident, well-structured artifact whose local sentences are all plausible and whose overall claim does not survive careful review. The reviewer-generator cost asymmetry that Borretti is complaining about is the same asymmetry every tech lead is now dealing with on code review.

The practical implications for senior engineers are concrete. Review time is now the bottleneck, not writing time. The old heuristics — 'this looks like code a competent engineer wrote, ship it' — are the exact heuristics the models have learned to spoof. If you are still using surface signals (naming, structure, comment quality, idiomatic style) as a proxy for correctness, you are going to miss subtle bugs because surface quality is now effectively free. The new signals you need are the ones that were always *harder* to fake: does the test actually exercise the change, does the diff match the stated intent, does the author answer follow-up questions in ways that reveal they understand what they did.

The second implication is cultural. Teams that respond to AI-assisted contributions by rubber-stamping them because they 'look fine' are quietly degrading their code quality in a way that won't show up for months. The failure mode is slow accumulation of plausible-but-wrong changes — the codebase equivalent of the Riemann 'proof.' You will not notice until an incident forces you to read something carefully, and then you will discover that three successive layers of abstraction are each lying to you in slightly different ways.

The third implication is about writing itself. If you are producing technical documents — design docs, post-mortems, RFCs — and you let a model do the final pass for 'polish,' you are moving your own output toward the same homogenized register that makes LLM-written math suspicious. The signal of individual voice, of specific hard-won opinion, is becoming more valuable precisely because it is harder to fake. Keep it.

Looking ahead

Borretti's piece will probably not be the last 'Re: [famous problem]' takedown this year, because the incentive to publish 'LLM makes breakthrough' is high and the cost of producing the artifact is essentially zero. Expect this to become a recurring genre, and expect the signal-to-noise ratio in the open-knowledge commons to keep degrading until the social cost of posting slop catches up with the technical cost of producing it. In the meantime, the practical defense is the same one mathematicians have used for three centuries and one that software engineers should adopt more seriously: trust proofs you have checked, trust results that have been independently reproduced, and treat fluent prose as the beginning of evaluation, not the end of it.

Hacker News 183 pts 176 comments

Re: Riemann

→ read on Hacker News
Kuyawa · Hacker News

I could not have solved GRH with Claude, openAI, DeepSeek or any other AI. Neither could 99.99% of you. AI is a tool, like a calculator that helped mathematicians long time ago calculate sine and cosine faster than we could, then computers, then the internet, now AI. Somebody has to use the tool and

doctorzook · Hacker News

If the measure of man is his ability so prove theorems, then the machines clearly outstripped most of us already. But by that measure, most of us had already been beaten by other men.I'll never be the next Euler or von Neumann or Einstein, and I'll confess to occasional pangs of regret tha

jareklupinski · Hacker News

> Why should we explore, discover, invent, succeed, become more, struggle, win, become divine? We don’t want to be more. We don’t want anything.why did i buy a skateboard? why did i break my arm riding it? why cant i wait to do it again?same reason we're in tar pits :) there's something

lf88 · Hacker News

This post resonates a lot with me. It's depressing to see so many people rooting for a future in which human intelligence becomes irrelevant.

fwlr · Hacker News

He’s right, it does read like a suicide note. It is precisely the tone of a terminal patient praising the skill, the care, the infinite patience of the doctor euthanizing him.That’s affected me quite heavily. I think I should probably spend the day offline, actually.

// share this

// get daily digest

Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.