Does the inability to perform simple literature/web searches really
reflect "superhuman intelligence"?

Steve Miller
Yeshiva University
August 3, 2026

I must say I had no interest in being dragged into such a
discussion, until a friend showed me OpenAI's stunning "Ten
advances in mathematics and theoretical computer science"
(https://openai.com/index/ten-advances-in-mathematics/) this
morning. Despite its nod to the Leiden Declaration and the role of
appropriate scholarly citations, it became quickly clear to me that
OpenAI seriously underperforms (to put it diplomatically) in this
arena. Though other controversies about OpenAI's alleged plagiarism
of works have been well-covered by the media, I felt this
mathematical angle deserves further scrutiny. I've tried to
communicate this to OpenAI, but as there was no contact email given
for this work, I've had to resort to ad hoc methods -- as of yet,
I've heard no fruitful response from those involved.

I'm deeply impressed by the first accomplishment, breaking the
Kabatiansky-Levenshtein bound. This surely demonstrates the
incredible intellectual and reasoning power of OpenAI's engines.
Their document
https://cdn.openai.com/pdf/reasoning-walkthroughs.pdf ("How the
ideas came together") is scientifically crucial because it explains
the key steps in the discovery process. Though mathematics often
operates in the currency of unsolved problems, it is usually the
underlying ideas behind breakthroughs -- and the assimilation of
those ideas by others -- which have the true value.

It didn't take long for me to notice multiple issues. Already on
p.1 I felt that the work of Henry Cohn and collaborators should
have been emphasized earlier up, to make clear which ideas are
taken from their work and which ideas are new.

The introduction of the Mellin transform to Cohn-Elkies functions
near the top of p.2 jumped out at me even more: although it is
presented as the first new innovation of the present work, it in
fact already appeared prominently in Section 5 of
https://arxiv.org/pdf/1603.04759. This decade-old arXiv reference
is not cited by OpenAI, but was surely used in training its models.
What other explanation is there for the similarity between these
passages? The literature on Cohn-Elkies functions is after all not
so large that prior work could have been overlooked.

Having barely progressed through the second page, I didn't bother
looking further -- it's reasonable to expect this problem occurs
repeatedly to other researchers, not just in the rest of the
document, but also in forthcoming other works. That this occurs
systematically is a problem.

I don't think it should fall on me to have to justify the academic
community's approach to literature citations. It obviously has many
benefits and exists for strong reasons, reasons which must
withstand corporate pressure. An academic who claims a stunning new
result, while plagiarizing less-critical details, both benefits
from the recognition of their accomplishment and yet is also
saddled with the embarrassment of representing other scholars'
ideas as their own. The two come together as a package.

Where the lines of responsibility between the actions of a
corporation and its individual domain experts is also an
interesting question, one not without precedent in the history of
Chemical Engineering (cf. also Matteo Wong's Atlantic piece "What
If We Held ChatGPT to the Same Standard as Claudine Gay?").

So that leads me to ponder how this problem could happen in the
first place: after all, a Google search yesterday morning for "Cohn
Elkies Mellin Transform" returned arXiv:1603.04759 as its first hit
(it no longer is: an hour later it was overtaken...by the new
OpenAI posting itself!). I can't answer this question, but several
possibilities come to mind, including that

1) OpenAI was not careful about appropriate citations,

or that

2) perhaps because the existing literature has been so thoroughly
digested into an LLM, OpenAI is not capable of finding appropriate
citations.

The former is sadly not inconsistent with OpenAI's legal history
with plagiarism allegations, so I'd like to think it's the latter.
However, the main financial incentive for OpenAI is presumably to
demonstrate the power of its products in an extreme case, the idea
being they'd be valuable to customers in more pedestrian settings.
Do they really fail at something as simple as citation sourcing?



Addendum (August 16, 2026). Since I've been asked by many
colleagues, I should mention I am still yet to receive a response
from OpenAI.