OpenAI recently claimed to have solved the Navier-Stokes Millennium Prize Problem. The announcement caught my attention while I was building my investigation agent Wargs from scratch - an agent that hypothesizes about a user query, asks follow-up questions, gathers evidence, contradicts, proposes alternate hypotheses, and finally reports.

I was looking for example queries a user might ask my agent, and the philosophical question of who gets the credit if an LLM/AI model solves a big math problem seemed like a good one to explore.

I didn’t have a good answer on the attribution topic yet, but let me put my bias out there first: I was (and am) vehemently against the idea of AI getting the credit for discoveries. The scene from Harry Potter comes to mind, where Harry kept praying to the Sorting Hat, “Not Slytherin”!

This post examines the attribution of credit when an AI model solves a significant mathematical problem and is based on the conclusions presented to me by Wargs and my interpretation of it. My findings reveal a stark divide between academic norms and corporate claims. Academic institutions and journals, guided by the 2026 Leiden Declaration on Artificial Intelligence and Mathematics and International Mathematical Union (IMU) endorsements, strictly maintain that authorship and credit belong exclusively to humans, as AI lacks the capacity for legal and professional responsibility.

While independent researchers often claim credit by treating AI as a sophisticated tool, corporate AI labs have attempted to claim breakthroughs as milestones of their model’s emergent properties - to win market and stakeholder trust. Ultimately, the ‘credit’ is bifurcated: academic prestige remains human-centric, while corporate ‘discovery’ is attributed to the developers of the underlying architecture.

Several examples have already played out in front of us that illustrate the different attribution models have worked out so far. Here I note a few:

The Prompter

Credit attributed to the human user/prompter, when the solution was reached through iterative prompting, strategic guidance, or the formulation of a novel approach that the AI merely executed.

For example, mathematician Ernest Ryu used GPT-5 to explore ideas but maintained a classical publishing style, taking full responsibility for the verification. Similarly, amateur Liam Price used ChatGPT to find a key insight for a 60-year-old problem, but the credit was tied to the experts who sifted through the raw output to distill a formal proof. The Leiden Declaration reinforces this by stating that responsibility for correctness remains exclusively with human authors.

There is, however, some uncertainty or tension as to whether the credit should go to the developers, especially when the discovery is an emergent property of a high-compute infrastructure inaccessible to independent researchers. This is evidenced by corporate labs claiming ‘milestones’ such as AGI (via obtaining 99% on ARC-AGI3) for internal models (or a combination with harnesses).

The Developer

Credit will be attributed to the developers of the AI model if the discovery was an emergent property of the model’s architecture and training data, rather than a result of specific user input.

This hypothesis is supported by corporate behavior but contested by academic norms. The Navier-Stokes dispute illustrates a corporate lab (OpenAI) attempting to claim credit for a result produced by an internal model, even when (allegedly) based on user-provided ideas. The mathematical community, however, views such claims as ‘corporate market timelines’ rather than academic authorship.

The ‘emergent property’ logic used by corporations is systematically rejected by academic gatekeepers, who frame authorship as a function of responsibility and accountability - something developers cannot assume for a model’s output.

The AI

Current legal and academic frameworks will refuse to grant ‘credit’ (in terms of authorship or legal ownership) to the AI itself, as AI lacks legal personhood and intent.

Multiple journal guidelines (e.g., Advances in Mathematics, MASE) explicitly forbid listing AI as an author or co-author because authorship implies responsibilities that only humans can perform. The Leiden Declaration explicitly ‘affirms the humanity of authorship,’ stating that credit should not be given to automated systems.

The Hybrid

The credit will be distributed as a ‘Collaborative Hybrid’ model, where the AI is cited as a co-contributor or a specialized tool, but the human remains the ‘responsible party’.

While papers often include ‘Contribution Statements’ or ‘Tool Disclosures,’ these do not grant the AI ‘credit’ for the work, but rather AI is treated as a disclosed tool, similar to a calculator. AI is thereby relegated to a disclosure statement rather than taking a 50-50 share for its ‘contribution’.

Academia vs AI-lab

To conclude, attributing credit for AI-solved math problem(s) is currently split by venue.

  • In academia, the human who prompts the AI and verifies the proof receives the credit, while the AI is disclosed as a tool.
  • In corporate contexts, the developers claim the discovery as a technological milestone.
  • AI itself is universally denied authorship due to its inability to take responsibility for the work. At least for now.

References