A Second Mathematician Accuses OpenAI of Hiding Its Training Data Sources

What's being alleged

According to The Verge, a second mathematician has come forward to accuse OpenAI of unethical and "dishonest" behavior, alleging the company lacks transparency about where the training data behind its models' mathematical capabilities actually comes from. The complaint lands just days after a separate, contentious dispute over whether OpenAI's models had benefited from unpublished mathematical work — research that was never formally released for anyone, human or machine, to read.

Why impressive math is becoming a liability

OpenAI and other labs have spent the last few years pointing to model performance on hard math benchmarks and open problems as evidence of genuine reasoning ability. That framing cuts both ways: the better a model looks on novel mathematics, the more working mathematicians start asking where, exactly, it learned that. If the honest answer traces back to a preprint shared privately, a conference talk, or an unpublished proof passed between collaborators, "we don't disclose training data" stops reading as boilerplate and starts reading as evasion.

The structural problem: no one outside the lab can check

The core issue in both disputes isn't a smoking gun — it's that there isn't one, because model builders don't publish training-data manifests. Researchers are left arguing from circumstantial evidence: a model producing reasoning that looks close to their own unpublished work, at a point before that work was public anywhere. OpenAI can deny it; the researcher can't independently verify it either way. That asymmetry is not a stable equilibrium — it's a trust deficit that compounds each time the same complaint resurfaces.

Why this matters beyond academia

If you're shipping a product on top of a foundation model — a trading signal generator, a code assistant, an underwriting model — you inherit the provenance of whatever that model was trained on, audited or not. You can't independently verify a frontier lab's data pipeline any more than a mathematician disputing a proof's origin can. That's a real, if diffuse, risk: reputational exposure if a provider's data practices are later found improper, and a preview of the kind of scrutiny that will eventually reach financial and code-generation use cases, where "where did this pattern come from" is not just an academic question.

What would actually resolve disputes like this

• Training-data manifests or provenance statements published alongside major model releases, even at a coarse level
• Independent third-party audits with a real ability to flag disputed inputs before a model ships
• A standing process for domain experts to raise "I believe my unpublished work is in here" claims and get an actual answer instead of silence

The pattern worth watching

One provenance dispute is a single story. Two in short succession, from separate mathematicians, in the same domain, is a pattern — and patterns are what move regulators and enterprise buyers, not isolated complaints. Expect this to keep surfacing as more domain experts compare their own unpublished reasoning against model outputs, and as AI labs face growing pressure to say more about where their training data actually comes from.