OpenAI Forms Math Advisory Group As AI Solves 100+ Open Problems Without Power to Slow Research

TL;DR

• OpenAI has created an external Math Advisory Group to guide evaluation and responsible disclosure after its internal models reportedly solved more than 100 open mathematical problems.
• The group is strictly advisory and has no authority to pause, veto, or redirect OpenAI's ongoing math research and model development.
• The milestone signals a shift from AI as a math assistant to AI as an autonomous discoverer, raising urgent questions about verification, credit, and the future of human-led mathematics.

A Breakthrough Hiding In Plain Sight

OpenAI's push into advanced mathematics has quietly crossed a historic threshold. According to details shared by the company this month, its latest AI systems have produced valid, verifiable solutions to more than 100 open mathematical problems spanning combinatorics, number theory, and related fields.

Unlike previous demonstrations centered on contest problems or formalizing known proofs, these are claimed to be novel results — problems with no previously known solution, many drawn from public lists and prize collections. OpenAI says each solution has been checked by human mathematicians and, in many cases, formalized for independent verification.

The company has not yet published the full list, but researchers familiar with the effort describe a pipeline where models propose conjectures, attempt proofs, critique their own work, and iterate for days or weeks with minimal human prompting. The scale, not just the difficulty, is what has stunned the math community.

Why OpenAI Created A Math Advisory Group Now

To help navigate the fallout, OpenAI has formed a new Math Advisory Group composed of outside academic mathematicians and experts in automated reasoning.

The group's mandate is focused on three areas: helping design fair benchmarks for mathematical discovery, advising on responsible publication and attribution when an AI solves a long-standing problem, and helping build better tools for formal verification and peer review.

OpenAI frames the move as a recognition that AI-driven discovery is now moving faster than traditional academic norms can handle. Who gets credit when a model solves a problem posed 50 years ago? How should journals handle a flood of machine-generated proofs? How do you prevent errors or duplicate claims? Those are the questions the advisors are expected to tackle.

Advisory In Name Only: No Brakes On Research

Crucially, the Math Advisory Group will not function as an oversight board.

OpenAI has confirmed the group has no governance power — it cannot pause experiments, veto releases, redirect research priorities, or slow down training of math-capable models. It does not have access to proprietary training data or internal safety reviews beyond what OpenAI chooses to share, and its recommendations are non-binding.

That distinction matters. Unlike an internal safety team or a corporate board, this is an external sounding board, not a control mechanism. OpenAI's math research, product integration, and scaling efforts will continue in parallel, regardless of any concerns raised by advisors.

Critics argue that structure limits the group's impact to ethics-washing, while supporters say it mirrors how other scientific advisory panels work — influence through expertise and public credibility, not hard power.

What 100+ Solved Problems Actually Means

Solving more than 100 open problems does not mean mathematics is solved. Most of the problems reportedly cracked are not Millennium Prize-level challenges like the Riemann Hypothesis or P vs NP, but mid-tier open problems — important to specialists, tractable enough for sustained machine search.

Even so, the implications are enormous. Historically, a strong mathematician might solve a handful of such problems in a career. An AI system doing so in bulk suggests a fundamental change in the economics of discovery: ideation and proof-search can now be scaled with compute.

Early reactions from mathematicians range from exhilaration to anxiety. Some describe using OpenAI's models as tireless collaborators that suggest lemmas and counterexamples humans would miss. Others worry about a deluge of correct but unilluminating proofs that advance problem counts without advancing understanding.

The Future Of AI-Driven Discovery

OpenAI's move points to a near future where AI labs operate like automated research institutes — generating conjectures, proving theorems, and submitting papers at machine speed.

For academia, that will require new infrastructure: faster formal verification systems like Lean, AI-assisted peer review, and clear rules for disclosure of machine assistance. For AI companies, it raises competitive stakes, with Google DeepMind, Anthropic, and open-source efforts racing toward the same goal of a generalist AI scientist.

The Math Advisory Group won't decide that race or slow it down. But it may help decide whether the resulting discoveries strengthen public mathematics or overwhelm it. If OpenAI's 100-problem claim holds up under independent scrutiny, September 2026 may be remembered as the month math changed from a purely human pursuit to a human-machine one.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.