AI benchmark maker under fire for hidden OpenAI funding

Imagine taking the SAT, only to find out your rival student secretly funded the test and got to flip through the questions beforehand. That is basically the trust crisis engulfing the AI community right now.

Reports highlighted mounting backlash against Epoch AI, a nonprofit research outfit primarily backed by Open Philanthropy. The issue? Epoch AI waited until December 20 to disclose that OpenAI had financially backed its expert-level math benchmark, FrontierMath. That benchmark happened to be the centerpiece OpenAI used to show off its flagship model, o3.

Even worse, OpenAI was granted privileged visibility into many problems and solutions long before the public found out.

Contributors kept completely in the dark

The delayed reveal sparked outrage from the very academics who wrote the test. On the LessWrong forum, a contractor posting as "Meemi" called out Epoch AI for non-transparent communication, arguing contributors should have known their work was being handed to OpenAI.

Stanford math PhD student Carina Hong took things further on X. She revealed that six key mathematicians behind the benchmark confirmed they had no idea OpenAI held exclusive access. Most said they probably would have walked away had they known.

Epoch AI co-founder Tamay Besiroglu admitted the organization made a mistake. He explained that contracts barred them from disclosing the tie-up until o3 launched, but acknowledged in hindsight that transparency should have been a non-negotiable term during deal negotiations.

A verbal agreement and unverified claims

To reassure skeptics, Besiroglu claimed FrontierMath's integrity remains intact. He noted that OpenAI agreed under a "verbal agreement" not to use the benchmark problems as training data to avoid teaching to the test. Epoch AI also keeps a separate holdout set to verify claims independently.

Yet that defense quickly hit another bump. Elliot Glazer, lead mathematician at Epoch AI, noted on Reddit that the organization hasn't actually verified OpenAI's o3 results yet.

Glazer said his personal take is that OpenAI's score is legit, but emphasized Epoch AI cannot officially vouch for the numbers until its own independent audit wraps up.

Why it matters

In our view, this mess exposes the messy conflict at the heart of AI benchmarking: everyone wants objective yardsticks, but almost nobody can fund them without Big Tech money.

Crafting PhD-level math problems requires elite brains and serious cash. But trading early access and non-disclosure clauses for funding makes an independent benchmark look like a bespoke marketing prop for OpenAI.

Relying on a "verbal agreement" to prevent data contamination is particularly baffling. When billions of dollars in enterprise reputation hang on model scorecards, handshakes simply do not cut it as a scientific safeguard.

For developers and buyers relying on these leaderboards to gauge raw capability, benchmark drama like this erodes confidence across the board. If non-profit evaluators cannot maintain absolute independence, third-party benchmarks risk becoming just another line in an AI lab's PR budget.