The boundary between artificial intelligence as a tool and AI as an independent scientific contributor has become increasingly blurred. In September 2024, OpenAI announced that an internal model trained since late August had reportedly resolved more than 100 long-standing open problems spanning most areas of mathematics—a claim that, if validated, would represent a watershed moment in both AI capability and mathematical research. The announcement arrives amid growing tension between AI companies seeking to demonstrate advanced reasoning capabilities and mathematicians concerned about the integrity of their field, raising fundamental questions about how machine-generated discoveries should be evaluated, communicated, and credited.

What OpenAI Actually Announced

According to the company's announcement, OpenAI's new internal model began training on August 28, 2026, and has since produced results across "most areas of mathematics." The company explicitly links this work to its claimed resolution of the Navier-Stokes Millennium Prize problem—one of seven problems carrying a $1 million prize from the Clay Mathematics Institute for a correct solution.

This represents a dramatic escalation from OpenAI's August disclosure of 10 mathematical and theoretical computer science results that either resolved or made substantial progress on open problems, some of which the company said had been formalized for machine verification. The sheer volume—more than 100 problems—distinguishes this announcement from previous AI contributions to mathematics, which have typically involved narrow applications or assistance with specific proof steps.

However, OpenAI has not publicly disclosed which specific problems have been solved beyond the Navier-Stokes reference, what methodologies the AI employed, or provided the proofs themselves for independent verification. This leaves the mathematical community and broader scientific establishment in a position of uncertainty about the substance and validity of these claims.

Establishing Oversight: The Math Advisory Group

Simultaneously with its research announcement, OpenAI established the Advisory Group on Mathematics and Artificial Intelligence, hosted at the prestigious Institute for Advanced Study in Princeton. The nine-member panel includes prominent mathematicians: François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood.

According to OpenAI, the advisory group will operate independently, with members able to provide unsolicited advice and make recommendations public. Critically, members will not be compensated by OpenAI—a structural choice designed to preserve independence. The group's mandate encompasses reviewing emerging results, assessing their significance, coordinating dissemination, and maintaining academic and professional standards in mathematical research. The group will also advise on how AI tools can support mathematical research and learning.

Notably, however, the group will not advise OpenAI on the pace of its internal mathematical research—a boundary that preserves the company's autonomy in capability development while seeking external validation for outputs.

The composition of the group is impressive—it includes Fields Medalists and leading researchers across multiple mathematical subdisciplines. The independence provisions are noteworthy: while members receive no compensation from OpenAI, they will necessarily depend on the company for access to the AI systems, the purported proofs, and the methodological details needed for meaningful evaluation. This structural dependency raises questions about how "independent" the oversight can truly be when the advisory group must rely entirely on OpenAI for the raw materials of their assessment.

The Broader Context: Mathematicians Push Back

OpenAI's announcement arrives against a backdrop of growing concern within the mathematical community about how AI companies are using mathematical research. The company's announcement notes that TechCrunch previously reported that 25 Fields Medalists—recipients of mathematics' highest honor—signed an open letter raising concerns about the use of mathematical research in AI development. While the specific concerns in that letter were not detailed in OpenAI's announcement, the timing suggests unease about AI companies treating unsolved mathematical problems as benchmarking datasets for model evaluation without adequate consultation with or benefit to the mathematical community.

This tension reflects what appears to be a fundamental asymmetry: AI companies possess computational resources and proprietary architectures that can generate mathematical results, but the mathematical community holds the expertise, standards, and institutional authority to validate those results. The formation of OpenAI's advisory group can be understood as an attempt to negotiate this asymmetry—a recognition that extraordinary claims require extraordinary validation from those with domain expertise. Whether this structure adequately addresses mathematicians' concerns about transparency, attribution, and the appropriate use of mathematical knowledge remains to be seen.

The Challenge of Validation and Opacity

The lack of immediate public disclosure of the specific problems solved, the proofs themselves, or detailed methodologies raises questions about validation that are particularly acute given the magnitude of what OpenAI asserts. The Navier-Stokes equations, which describe fluid motion, have resisted complete mathematical solution for over 160 years. A valid proof of their properties would not only claim a million-dollar prize but fundamentally advance our understanding of fluid dynamics with implications spanning weather prediction, aerodynamics, and ocean currents.

Traditional mathematical practice involves priority claims being accompanied by detailed exposition that allows peer review. The announcement of results before their full publication represents a departure from this norm, though OpenAI's establishment of the advisory group suggests recognition that external validation is necessary. This creates an unusual situation: the mathematical community is being asked to take seriously claims of extraordinary achievement while the evidence remains sequestered.

Mathematical proofs can be extraordinarily complex, and even human-generated proofs sometimes contain subtle errors that take years to identify. The 1993 announcement of Andrew Wiles's proof of Fermat's Last Theorem, for instance, was followed by the discovery of a gap that required additional work to resolve. OpenAI's mention that some previous work "had been formalized for machine verification" is significant, as formal verification systems can provide high confidence in correctness by checking proofs against rigorous logical foundations. However, the extent to which the 100+ claimed solutions have undergone such verification remains unclear from the announcement.

The advisory group's role in "assessing significance" addresses another important dimension beyond correctness. Not all open problems are equally valuable; some represent genuine barriers to progress in important areas, while others may be technical curiosities or problems that remain open primarily because few mathematicians have devoted sustained attention to them. The mathematical community's judgment about which problems matter and why represents expertise that complements technical validation.

Transparency, Reproducibility, and the Future of Mathematical Practice

The OpenAI announcement raises fundamental questions about transparency and reproducibility in an era of AI-generated mathematics. Traditional mathematical practice involves publishing proofs in sufficient detail that other mathematicians can verify them independently. This serves multiple purposes: it allows error detection, enables others to build on the work, provides pedagogical value as students and researchers learn from the methods employed, and ensures that mathematical knowledge becomes part of the collective intellectual commons.

The advisory group's mandate to coordinate dissemination suggests OpenAI recognizes the importance of communication to the broader mathematical community, but the details of how dissemination will work remain to be determined. Will the full proofs be published in traditional journals, subject to conventional peer review? Will they be made available in formats that allow independent verification? Will the AI systems themselves be made accessible to researchers who want to understand or extend the methods? These questions have no clear answers yet.

If AI systems generate proofs through processes that are partially opaque—either because the reasoning involves neural network computations that resist human interpretation, or because companies treat the systems as proprietary—it could create challenges for traditional mathematical validation and knowledge-building. A proof that can only be verified by running it through the same AI system that generated it, or that requires access to proprietary tools, would represent a significant departure from mathematical norms of transparency and reproducibility.

There is also the question of understanding versus verification. Even if a proof is formally correct, mathematicians value understanding why a result is true and what insights the proof provides. A machine-generated proof that is correct but inscrutable may be less valuable to the mathematical community than a human-readable proof that illuminates underlying structure. The advisory group's role may include assessing not just correctness but also the interpretability and pedagogical value of AI-generated results.

Comparing Perspectives: Enthusiasm and Caution

The available reporting on OpenAI's announcement reflects a range of perspectives on what this development means. Some coverage emphasizes the potential breakthrough nature of the achievement, framing it as evidence of AI systems reaching new levels of abstract reasoning capability. This perspective sees the advisory group as a responsible governance mechanism that allows rapid capability development while ensuring appropriate oversight.

Other perspectives, reflected in the context of the Fields Medalists' letter, emphasize concerns about process, transparency, and the relationship between AI companies and the mathematical community. From this viewpoint, announcing the solution of 100+ problems without providing the solutions themselves, or details about which problems were solved, represents a problematic departure from scientific norms. The advisory group, while potentially valuable, may be insufficient if it operates primarily as a validator of results generated through opaque processes rather than as a partner in shaping how AI is integrated into mathematical research.

These differing perspectives reflect broader debates about AI development: should we prioritize rapid capability advancement with governance mechanisms added to manage risks, or should we insist on transparency and community involvement from the outset? The mathematical community's response to OpenAI's approach may influence how similar questions are answered in other scientific domains.

Implications for Scientific Discovery

If the claimed resolution of 100+ open mathematical problems withstands scrutiny, it would demonstrate that AI systems have achieved a level of abstract reasoning capability that represents genuine scientific contribution beyond computational assistance or narrow problem-solving. This would have implications extending beyond mathematics to other fields where complex reasoning and problem-solving are central—from theoretical physics to computer science to philosophy.

However, significant uncertainties remain. The specific problems solved have not been publicly disclosed, making it impossible to assess whether they represent major breakthroughs or more incremental advances. The extent of formal verification is unclear. The advisory group's effectiveness in maintaining standards while operating independently of OpenAI's research pace remains to be demonstrated. And the mechanisms for disseminating results in ways that serve the broader mathematical community have yet to be established.

The path forward requires neither uncritical enthusiasm nor reflexive skepticism, but rather the development of robust validation processes, transparent disclosure of methods and results, and governance structures that preserve scientific integrity while enabling beneficial innovation. OpenAI's mathematical advisory group represents one experiment in this direction; its success or failure will offer valuable lessons for the broader integration of AI into scientific research.

What Comes Next

The mathematical community's response to these developments will likely shape how AI-generated scientific discoveries are received in other fields. If the advisory group successfully validates significant results while maintaining rigorous standards and ensuring appropriate dissemination, it could provide a model for other domains. If the process reveals fundamental challenges in validating AI-generated work, or if the results prove less substantial than initially claimed, it will highlight the need for more careful approaches before making extraordinary claims.

Several key questions will determine the trajectory of this development:

Validation timeline: How quickly will the advisory group be able to assess the claimed results? Will validation happen on a timeline that satisfies both the mathematical community's need for rigor and OpenAI's interest in demonstrating capability?

Disclosure practices: Will OpenAI publish full proofs in traditional venues, making them available for independent verification? Or will access remain mediated through the advisory group and the company itself?

Reproducibility: Will other researchers be able to reproduce or extend the AI-generated results? This depends on both the availability of the proofs themselves and access to the AI systems or methodologies used to generate them.

Impact on the field: If the results are validated, how will they influence mathematical research? Will they open new areas of inquiry, or primarily close off problems that were already somewhat isolated from mainstream mathematical development?

Governance effectiveness: Can the advisory group maintain genuine independence while depending on OpenAI for access to systems and results? Will its recommendations be followed, and will it have the authority to raise concerns publicly if necessary?

Ultimately, the integration of AI into mathematics and science more broadly will succeed only if it preserves the core values that make these enterprises trustworthy: commitment to truth, transparency that enables independent verification, and institutional structures that maintain standards. The coming months and years will reveal whether OpenAI's approach achieves this balance, or whether alternative models are needed to ensure that AI-generated scientific discoveries serve the broader goals of human knowledge and understanding.

The stakes extend beyond mathematics. If AI systems can genuinely contribute to solving long-standing problems in abstract reasoning, they may also be capable of advancing other areas of fundamental research. But the credibility of such contributions depends on establishing processes that the scientific community can trust—processes that balance innovation with integrity, speed with rigor, and proprietary development with open science. OpenAI's mathematical advisory group is an early test of whether such balance is achievable.

Correction, October 25, 2026: An earlier version of this article gave the year of the start of the model's training as 2024. The correct year is 2026.

Editor's note, October 25, 2026: We published a second article on this announcement on September 22, 2026, in error. It now redirects here.

Sources

  1. OpenAI Launches Math Advisory Group as AI Advances Mathematical Research With 100+ Solutions — itechpost.com
  2. Alexey Tulia at Tech Race Summit: Why AI Agents Raise the Stakes for Engineering Leaders — techtimes.com
  3. Why is Gravity's strength so hard to measure? NIST has a new clue — techexplorist.com
  4. Osteoclastogenesis with FN1-associated cell differentiation, immune communication, and ECM remodeling in male SAMP6 mice - Communications Biology — nature.com