Mathematician Andreas Thom has publicly accused OpenAI of unethical data practices regarding its recent breakthroughs in formal mathematics. Thom posted a series of updates on Mastodon arguing that the company incorporated unpublished work from his own research circle into its training sets. He claims this interaction with the ChatGPT chatbot prior to official announcements gave the model an unfair advantage, describing the situation as dishonest and lacking transparency. This follows a previous dispute where another researcher alleged similar misconduct concerning unpublished mathematical proofs.
The issue matters because it challenges the integrity of the entire field of AI-driven mathematics. If models learn from private, unpublished research rather than established public knowledge, the resulting proofs lack verifiable foundations. This creates significant uncertainty for mathematicians who need to trust that new discoveries are genuine breakthroughs and not the product of data contamination. The situation forces the academic community to question how training data is sourced and verified in high-stakes scientific domains.
- Thom specifically cites interactions with the ChatGPT bot before OpenAI’s public claims.
- Accusations focus on the inclusion of unpublished work within training datasets.
- Concerns centre on the inability to verify the provenance of new mathematical proofs.
