Recently the mathematical world was shaken when OpenAI released a solution to the Navier-Stokes problem, as supported by a valid, AI-generated proof. This brought a heavy question to the forefront of many people’s minds; if AI can make its own novel discoveries, then what purpose does human research have anymore? This article aims to discuss the context around OpenAI’s claims, and whether findings by AI can be truly original, or if they all boil down to combining previous findings in a new way.
Recently the mathematical world was shaken when OpenAI released a solution to the Navier-Stokes problem, as supported by a valid, AI-generated proof. This brought a heavy question to the forefront of many people’s minds; if AI can make its own novel discoveries, then what purpose does human research have anymore? This article aims to discuss the context around OpenAI’s claims, and whether findings by AI can be truly original, or if they all boil down to combining previous findings in a new way.
To figure out whether AI can be accredited for its discoveries, it is first worth considering how AIs – like OpenAI’s internal model which solved the Naiver-Stokes problem – function. Despite not stating which type of model they used, OpenAI specialises in Large Language Models (LLMs) and Large Reasoning Models (LRMs). Because of this, it is reasonable to believe that their internal model is a LRM of some form.
So, what is an LLM? A LLM is an AI model that predicts the next item in a pattern. This could be the next word in a sentence, the next line of code in a program, or the next number in a sequence. For example, the AI behind ChatGPT is an LLM. The reason why LLMs, like ChatGPT, are so convincing is that they are trained on so much data that they become very good at predicting patterns (arXiv: Do Reasoning Models Reason?, 2025). For example, this means that you could ask an LLM a physics question, and because it has processed so much data related to physics, it would be able to string together an answer. However, if the vast majority of the LLM’s data was incorrect, it would form an incorrect answer – this is because it has been trained on incorrect patterns.
Since a lot of data generated by humans is unreliable, companies like OpenAI created LRMs to produce more accurate answers. The issue is, LRMs are still built on the same underlying principles as LLMs. This means LRMs are still just predicting how a given pattern – like a maths question – should be answered from their set of analysed data. The difference between a LRM and LLM is that, whilst a LLM predicts and presents its first answer, a LRM attempts to verify its output beforehand. One way a LRM achieves this is by choosing the most common answer produced by multiple LLMs. Another way it attempts to do this is by having a Verifier AI review the LLM’s answer and then feedback to the LLM, prompting it to regenerate its prediction using the feedback. This process is then repeated until the answer is accurate enough to be displayed to the user (arXiv: Do Reasoning Models Reason?, 2025). What this means is that even the best LRMs are still just predicting the next item in a given pattern, just in a more roundabout way.
So how could AI solve the Naiver-Stokes problem without data to draw predictions from; who did the model copy? It just happens that, in New York University (NYU), Professor Tristan Buckmaster and Anthropic employee Levent Alpög were working towards a solution for the Naiver-Stokes problem simultaneously to OpenAI. To help streamline their work, they used OpenAI’s Codex, which meant that the findings they had supplied then became data that OpenAI could have used to train their models. This meant that, according to OpenAI, although “unlikely, we cannot rule out that [their data] helped improve our models” (BBC, 2026). Bearing in mind how LRMs function, and the fact that OpenAI’s internal model may have been trained on Buckmaster’s findings in NYU, it is not implausible to suggest that the model could have used their findings to form its final solution. Therefore, if it did, the AI’s findings were not truly original, but were rather an extension of a novel approach produced by humans (and not just those developed by Buckmaster; this would apply to all the data OpenAI accessed to create the final proof). This would suggest that AI has no claim to its research and is incapable of making something unique. If AI does not have data on the problem, it is unable to produce a meaningful prediction: here it had Buckmaster’s data, and so was able to extend it via predictions to procure a solution.
Closed case, right? The issue is OpenAI’s solution took a significantly different approach to the problem. If their AI was simply extrapolating from concurrent work, then it would have started in a similar place with similar assumptions. However, OpenAI claims that their proof, down to the exact results proven, vary greatly. In addition, they state that the AI and researchers alike did not see any of Buckmaster’s user data or concurrent work during their production of their Naiver-Stokes solution. OpenAI thus maintains that it was a coincidence that both groups – OpenAI and Buckmaster – came out with a solution (OpenAI, 2026). Although suspicious, if this was the case, does this mean that AI had made its own novel discovery?
That question would become a matter of deciding at what point extrapolating from a pattern shifts from plagiarism to unique work. Consider an artist; one may look at other pieces of art and become inspired by them, and then decide to combine a collection of styles and techniques into a new painting. These new styles, when put together, could be considered its own artistic approach. What the artist has done is combine a series of patterns into their own form of expression or unique idea. It could be argued that some AI systems combine patterns in new ways like this, since it has access to and processes more information than a human ever could. An example of this is Google’s AlphaEvolve, which contributed to research concerning numerous maths problems, including Strassen’s 1969 algorithm which is used for matrix multiplication (Google DeepMind, 2025). AlphaEvolve was able to combine maths in its training data on a scale inaccessible to humans to form new discoveries – discoveries which could be considered AI-owned.
Overall, it is hard to say whether AI holds ownership over its findings. Ultimately, all LRMs stem from combining human data in new ways, but the question as to when those combinations become subtle enough that they can be considered a novel discovery is a difficult one to answer. At this point in time, I am inclined to question OpenAI’s statements about Buckmaster’s work, as I do not believe AI is powerful enough to produce anything more than extrapolations and combinations of human data. Thus, I am currently of the opinion that OpenAI’s findings are not their own. However, as AI developed behind closed doors – like OpenAI’s internal model used for the Naiver-Stokes problem – becomes more powerful, the patterns and predictions it can see in training data may become more subtle to the point they could be considered an AI-driven discovery. If we are not careful and leave companies like OpenAI unchecked, one day AI may just replace us completely; from manual labour to research roles, there will not be anything intellectual left for people to enjoy.