News · 2026-08-10

The viral Riemann result an AI supposedly proved is not in the literature

A claim that circulated widely on Reddit and social media -- that an AI model raised the proven fraction of Riemann zeta zeros lying on the critical line from about 41.6 percent to 67.2 percent -- does not correspond to any result in the published literature. The closest primary source, a November 2025 expository note by Daniel Goldston and Ade Irma Suriajaya, says a two-thirds result would follow only if a standing assumption could be removed from an existing argument, and states that this has not been achieved. No proof-assistant artifact for the claimed result exists.

Key facts

• The circulating claim: a jump from roughly 41.6 percent to 67.2 percent of zeta zeros proven to lie on the critical line, credited to an AI model.
• The Goldston and Suriajaya note says the two-thirds figure follows only if the Riemann Hypothesis can be removed from Montgomery's simple-zero argument, and that this is not yet obtained unconditionally.
• The note's 67.92 and 70.37 percent figures refer to simple zeros assuming the hypothesis -- a different statement entirely.
• No Lean or Coq formalization of the claimed result surfaced on any of the pages checked.

Here is why the confusion is easy to fall into. The Riemann Hypothesis is the conjecture that every non-trivial zero of the zeta function sits on one particular vertical line in the complex plane. Nobody has proved it. What mathematicians have done for a century is prove partial results: statements of the form "at least this fraction of the zeros provably lie on the line." That fraction has crept upward over decades of hard work, and the benchmark result in the modern literature is a bound in the low forties, established in a 2018 paper by Pratt, Robles, Zaharescu, and Zeindler. Moving it to two-thirds in one step would be one of the biggest results in analytic number theory in fifty years.

The trap is that percentages near 67 and 70 genuinely do appear in this literature -- but attached to a different question. There are results about simple zeros, meaning zeros that are not repeated, and there are results that hold conditionally, meaning they assume the Riemann Hypothesis is true and derive consequences. "Two-thirds of zeros are simple, assuming the hypothesis" and "two-thirds of zeros provably lie on the line, assuming nothing" are both sentences containing two-thirds, and only one of them would be news. Assuming the thing you are trying to prove is not a small footnote. It is the entire difficulty.

The same wave carried two other claims worth correcting. One held that the "Imbalance Conjecture" had been proven. There is a real recent theorem -- How Label Imbalance Shapes Geometry, dated May 3, 2026 -- that resolves the multiplicity-one imbalance conjecture from a 2024 paper on multi-label neural collapse. That is a genuine contribution with a new spectral-control framework behind it, and it is a specific theorem about the geometry of neural network representations, not a broad conjecture in mathematics. The other claim held that Teschner's bondage-number conjecture had been disproven. The relevant 2012 paper by Gagarin and Zverovich does the opposite: it settles Teschner's conjecture affirmatively for almost all graphs, using computer algebra for some finite checks. Disproven and settled-affirmatively are not close.

What all three share is a failure mode worth naming, because it will keep happening. None of these claims was invented from nothing. Each attached itself to a real paper and then drifted: a conditional result lost its condition, a narrow theorem lost its qualifiers, an affirmative settlement flipped sign. The drift happened in the retelling, not in the source. And the AI framing made each one more shareable, because "an AI proved X" is a much better headline than "a 2012 graph theory paper settles a conjecture for almost all graphs."

The check that would settle any of this is available and nobody ran it. A machine-checkable proof in a system like Lean or Coq either compiles or it does not, and there is no room for a conditional result to quietly shed its condition -- see our lesson on what a proof assistant is. No formalization surfaced for any of the three claims. That absence is not proof the claims are false, but for a result of this magnitude the formalization is exactly what a serious announcement would lead with.

The honest caveat: AI systems are contributing real mathematics, and dismissing every such claim would be its own error. The point is that the verification bar for "an AI proved a major theorem" is a preprint, a named author, and ideally a formal proof -- not a screenshot. This is the same shape as the 70 percent circular AI revenue figure: a specific number, confidently repeated, that nobody could trace to a source.

Primary source, verified: read the paper → (arXiv 2511.20059)

Ground Truth. (2026, August 10). The viral Riemann result an AI supposedly proved is not in the literature. Ground Truth. https://groundtruth.day/news/the-viral-riemann-bound-does-not-appear-in-the-literature.html

@misc{groundtruth:the-viral-riemann-bound-does-not-appear-in-the-literature,
  title  = {The viral Riemann result an AI supposedly proved is not in the literature},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/the-viral-riemann-bound-does-not-appear-in-the-literature.html}
}

Topics: mathematics · verification · research-integrity · ai-claims · proof-assistants