OpenAI's AI-generated math manuscripts spark academic debate

A massive output of math

OpenAI recently released more than 700 manuscripts covering over 370 different mathematical topics. These works were produced using a model that has not yet been publicly released. The company reported that the AI was able to generate these manuscripts in an average of just three hours each.

The scope of these publications includes some of the most famous unsolved problems in the field. For instance, the manuscripts touch upon the Riemann hypothesis, a complex problem involving prime numbers that has puzzled experts for a long time. This follows similar efforts by other companies, such as Anthropic, to use AI for mathematical exploration.

While the AI has not officially solved these major problems, the results are being framed as significant progress. However, the sheer volume of the release has sparked a heated debate among academics. Critics are questioning whether these outputs constitute genuine scientific achievement or merely a demonstration of computing power.

The value of the process

A primary criticism from the academic community centers on the difference between finding an answer and understanding a method. The German Mathematicians' Association (DMV) argues that mathematics is about the path taken to reach a conclusion. They suggest that AI is currently results-oriented, focusing on the final solution rather than the development of new methods.

Demian Nahuel Goos, a mathematician, noted that the proof itself provides very little insight. He argues that the goal of research is to discover how a result is achieved. If humans cannot verify the logic behind an automated result, the output may be considered virtually worthless in a scientific context.

Furthermore, there are concerns regarding scientific integrity and the rules of attribution. Goos stated that failing to follow established rules for citing others' work is unscientific. This is particularly relevant when AI models are trained on existing data without clearly identifying the original human contributors.

Risks of unverified data

The DMV also highlighted a technical risk regarding the data used to train these models. AI models often utilize existing knowledge that has not been fully vetted or verified by experts in specific subfields. This means that the proofs generated by AI might rely on assumptions that are not yet proven to be correct.

This issue creates a potential cycle of misinformation within the mathematical community. If unverified information is integrated into new proofs, it could lead to errors that are difficult to trace. Experts warn that such results should be treated with extreme caution because the underlying knowledge is not yet secured.

Some researchers have already expressed concern about the impact on open exchange. Jürg Kramer, president of the DMV, noted that some researchers may stop posting their findings online. He fears this will slow down progress by preventing the open sharing of ideas that has historically been central to the field.

Redefining the PhD

The rise of AI in mathematics is also putting pressure on the training of future scientists. Christian Hesse, a professor of mathematics, suggested that the traditional doctoral thesis may need to be reinvented. If an AI can solve a problem in minutes that would take a human years, the focus of a PhD must shift toward other skills.

Hesse believes that future research should place a higher value on the development of new questions and critical thinking. Instead of just finding a correct answer, students will need to demonstrate a deep understanding of the underlying concepts. The goal is to move away from the illusion of human irrereplaceability in simple calculation.

Despite these concerns, some see the potential for AI to discover connections that humans might never notice. Hesse suggested that machines might eventually teach us what mathematical creativity actually looks like. He notes that while it may be frustrating for humans who spent years on a single problem, the technology is still in its early stages.

The future of verification

There is an ongoing debate about how the scientific community should handle these AI outputs. One possibility suggested by the DMV is that traditional publishers could play a larger role by only releasing works that pass a rigorous human peer-review process. This would create a barrier for AI-generated content that hasn't been checked by experts.

However, others argue that closing off information behind paywalls or restricted access is counterproductive. They believe that science thrives on the free exchange of ideas rather than exclusion. The challenge lies in finding a balance between encouraging AI-assisted progress and maintaining the standards of human-led verification.

OpenAI maintains that their goal is to expand the boundaries of human knowledge and provide tools to advance these fields. They intend to evaluate their internal models to develop faster tools for researchers. Ultimately, the revolution may be a shift in how we define mathematical understanding, moving from the result to the insight.