AI Benchmarking Group Faces Backlash for Delayed Disclosure of OpenAI Funding

In recent news surrounding AI development, allegations of impropriety have emerged regarding AI math benchmarks created by Epoch AI, a nonprofit organization. The controversy stems from the revelation that OpenAI provided funding for the FrontierMath benchmark, a tool designed to assess the mathematical capabilities of AI systems. This announcement has raised questions about transparency and integrity within the AI community.

Background on FrontierMath and OpenAI’s Involvement

On December 20, Epoch AI disclosed that it had received support from OpenAI for the development of FrontierMath. This benchmark includes expert-level problems meant to evaluate AI’s mathematical skills and was utilized by OpenAI to showcase its upcoming flagship AI, known as o3.

Concerns Raised by Contributors

A contractor for Epoch AI, identified as “Meemi” on the forum LessWrong, expressed concerns about the lack of transparency regarding OpenAI’s involvement. According to Meemi:

• The contributors to FrontierMath were not informed about OpenAI’s funding until it was publicly announced.
• There should have been clearer communication about how this funding could affect their work.

Impact on the Reputation of FrontierMath

On social media, users have voiced worries that the undisclosed funding might undermine the credibility of FrontierMath as an impartial benchmark. Critics point out that OpenAI had access to various problems and solutions within the benchmark, which was not divulged until the December announcement.

Allegations of Exclusive Access

Stanford PhD student Carina Hong reported that some mathematicians involved in FrontierMath were unaware of OpenAI’s exclusive access to the benchmark. She noted:

• Six mathematicians confirmed they would have reconsidered their contributions had they known about OpenAI’s privileged access.

Epoch AI’s Response to Transparency Issues

In response to these allegations, Tamay Besiroglu, associate director and co-founder of Epoch AI, acknowledged that the organization should have been more transparent. He stated:

• Epoch AI was limited in its ability to disclose the partnership until the launch of o3.
• The organization should have prioritized transparency regarding who had access to the benchmark contributors’ work.

Assurances Regarding Data Use

Despite the concerns, Besiroglu emphasized that OpenAI has a verbal agreement with Epoch AI not to use the FrontierMath problem set for training its AI. He added:

• Epoch AI maintains a separate “holdout set” to ensure independent verification of FrontierMath results.
• OpenAI supports this decision to keep a separate, unseen holdout set for validation purposes.

Ongoing Verification Challenges

However, the lead mathematician at Epoch AI, Ellot Glazer, mentioned on Reddit that the organization has yet to independently verify OpenAI’s results from FrontierMath. He noted the complexity of the situation:

• While he believes OpenAI’s score is legitimate, independent verification is necessary.

The Broader Implications for AI Benchmarking

This situation underscores the challenges faced in creating empirical benchmarks for AI evaluation. It raises important questions about securing funding for benchmark development while avoiding conflicts of interest.

For more information on AI benchmarks, visit this link. To learn about OpenAI’s advancements, check their official page here.