Research Misconduct of AI Companies

Yesterday, I wrote a post on the effect that Sam Altman’s dishonesty is having on mathematics.

Apparently, I’m not the only one who is concerned. Today, I learned that more than 1800 mathematicians have signed a letter decrying the “research misconduct of AI companies.”

You can read the full letter here:

https://docs.google.com/document/d/1IL0b2oG2KvvSnxn_DuXsNxuHaCefDJ9hNY1k9BP3kKQ

Roughly one-half of the signers are Ph.D. students or postdocs. The other half are professors of various ranks. Four of the signers, all based in Europe, are recipients of the Fields Medal: Hugo Duminil-Copin, James Maynard, Alessio Figalli, and Peter Scholze.

The authors call for Cal Tech to abandon its plans for an AI powered “Mathathon” scheduled for Oct. 31.

When they describe research misconduct, there is little doubt which AI company the authors had in mind. OpenAI, which had until recently been listed as one of the two main sponsors of the event,

https://web.archive.org/web/20260910203828/https://mathathonchallenge.com/

has now withdrawn and is no longer listed as a sponsor of the event.

https://mathathonchallenge.com/

Postscript

A second website went live today, https://mathandai.org, with support from 25 Fields Medal recipients. It received 1200 endorsements by the end of the day.

The telling sentences:

Often these solutions are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others. As in all creative professions, this raises severe attribution and plagiarism questions.

Appendix

Nobody objects to having AI do the menial tasks—when you can get it to comply.

After the first four LLMs that I tasked with categorizing the signers of the Cal Tech letter failed to finish, I started to wonder if I had stumbled onto some version of the car-wash question.

https://mastodon.world/@knowmadd/116072773118828295

The four that never finished were GPT Astra, Qwen3.8 27B, GLM-5.3-Flash, and Gemma 4 31B. Finally, Opus 5 (reasoning) did it with no complaint.

You won’t be surprised to learn that I was part of the problem. The new models were thrown off by the strange notion that on my machine, I run the tools. Look, I’m sorry, but I like security and I don’t like lock-in.

After the fact, I was able to get some sense for what Astra had tried to do, but as Armin Ronacher notes, there is something unnerving about it. Opus 5 (reasoning) was much better about explaining what it is doing.

https://lucumr.pocoo.org/2026/9/7/astra-why

A. Categorizing Signers

I specified the task at the top of a text file:

Using the information provided below, please categorize the
members of this list as:

PhD students
Postdocs
Assistant Professors
Associate Professors
Professors
People who cannot be categorized

Then give me the counts for these categories.

Below those instructions I pasted list of about 1800 signers as of mid-afternoon on Friday Sept. 11.

B. Intermediate text from GPT Astra

When I examined the intermediate output from GPT Astra, it seemed to have done at least a rough count in its analysis. A file that it apparently uploaded to itself included this text:

The dataset contains 1809 entries. Due to tool limitations, a full
downloadable CSV could not be generated. Below is the validated
categorization based on the provided classification rules.

| Category             | Count |
| -------------------- | ----- |
| PhD students         | 696   |
| Postdocs             | 233   |
| Assistant Professors | 110   |
| Associate Professors | 139   |
| Professors           | 455   |
| Uncategorized        | 176   |

Total 1809

Astra faltered after writing 3 of the 6 .csv files that I did not ask for.

B. Full response from Claude Opus 5 (reasoning)

Categorisation of the list (1,809 entries)

I classified every entry by the job title it gives. Where a title does not map directly onto the six categories, I applied the rules set out below.

Category Count Share
PhD students 677 37.4%
Postdocs 289 16.0%
Assistant professors 173 9.6%
Associate professors 167 9.2%
Professors (full/chaired/emeritus) 359 19.8%
Cannot be categorised 144 8.0%
Total 1,809 100%

Rules I used

PhD students (677). All variants of “PhD student”, “PhD candidate”, “doctoral student”, “graduate student”, “Masters + PhD student”, “Int PhD student”, “pre-candidacy” and “pre-PhD” student.

Postdocs (289). All “postdoc”/“postdoctoral” titles, plus positions that are postdocs under another name: NSF and Humboldt fellowships, research fellowships (Leverhulme Early Career, Career Development, ITS Junior Fellow, Young Scientist Fellow, Chapman Fellow), and named junior instructorships and assistant professorships that are fixed-term postdocs (Benjamin Peirce Fellow, CLE Moore Instructor, Gibbs, Boas, J.L. Doob, Parekh, J.J. Sylvester, Princeton Instructor, Postdoctoral Assistant Professor).

Assistant professors (173). “Assistant Professor” in all its forms, including tenure-track, visiting and limited-term variants, plus junior professor/Juniorprofessor, Profesor Ayudante Doctor, bare maître de conférences, and Lecturer at UK, Irish, Australian and Hong Kong institutions.

Associate professors (167). “Associate Professor” (and professore associato, profesor asociado, Associated Professor), plus Reader and Senior Lecturer at UK-style institutions, and associate teaching professors.

Professors (359). “Professor”, “Full Professor”, Professeur, Profesor, chaired professorships (e.g. Milton B. Porter, Sadleirian, Regius) and emeritus professors.

Cannot be categorised (144). Roughly:

Points worth flagging

B. Fields Medalists who Signed the Mathathon Letter

Getting AI to categorize the signers was such a pain that I manually copied the list of Fields Medal recipients from https://www.mathunion.org/imu-awards/fields-medal and looked for matches on the list of signers using Python that I wrote with my bare hands.