Epoch AI, a nonprofit developing mathematical benchmarks for AI, has faced criticism for not disclosing funding from OpenAI until December 20, sparking concerns in the AI community.
The organization revealed that OpenAI supported the creation of FrontierMath, a test designed to measure AI’s mathematical capabilities. FrontierMath was also used by OpenAI to showcase its upcoming flagship AI model, o3.
However, contributors to the benchmark were reportedly unaware of OpenAI’s involvement until the announcement.
On the forum LessWrong, a contractor for Epoch AI, using the username “Meemi,” expressed dissatisfaction with the lack of transparency, stating that contributors should have been informed about OpenAI’s role.
Concerns were also raised on social media, with Stanford PhD student Carina Hong claiming that OpenAI’s access to FrontierMath was privileged and unknown to many mathematicians involved in the project.
Some contributors reportedly questioned whether they would have participated had they known about OpenAI’s involvement.
Epoch AI’s associate director, Tamay Besiroglu, admitted the organization should have been more transparent.
He explained that contractual restrictions prevented earlier disclosure but acknowledged it was a mistake not to insist on transparency from the start. Besiroglu added that OpenAI has verbally agreed not to use FrontierMath’s problem set for AI training and highlighted the presence of an unseen holdout set for independent verification.
Despite these assurances, doubts remain. Epoch AI’s lead mathematician, Elliot Glazer, noted that they have not independently verified OpenAI’s results on FrontierMath but stated his belief that the results are legitimate.
This controversy highlights the challenges of developing unbiased AI benchmarks while relying on funding that could create perceptions of conflicts of interest.
As AI technologies advance, maintaining transparency and objectivity in benchmark development will be crucial to ensure credibility and trust within the industry.
