Mathematician Andreas Thom has criticized OpenAI's opaque approach, suggesting that the company's AI models may have utilized user interaction data during the process of achieving results in the field of mathematics.

In a post on Mastodon, Thom pointed out that one of the mathematical achievements recently announced by OpenAI involves "non-sofic groups" (infinite mathematical structures that cannot be approximated by finite structures), which is his area of expertise. While OpenAI has acknowledged extensive use of previous research by Thom and his colleague Gabor Kun, Thom stated that he has not received a satisfactory explanation from OpenAI regarding whether interactions with ChatGPT were included in the training data or the inference process.

While OpenAI denied directly accessing specific user data to solve the problem, it did not completely rule out the possibility that anonymized usage data contributed to model improvements. Thom argues that such a response is disingenuous, as the intellectual content of mathematical ideas remains even if names are removed through anonymization.

Following this incident, concerns have been raised within the mathematical community regarding the risk of triggering development races among tech companies before research results are officially published.

Source: