OpenAI floods math with 719 AI papers; Korea hails speed, flags clarity strain - CHOSUNBIZ

As OpenAI released 722 math paper drafts written using artificial intelligence (AI), Korean mathematicians said AI excels at rapidly drawing on existing research and consolidating knowledge across different fields. But some noted that parts of the AI-written papers are difficult to follow in structure and notation, which could mean it takes considerable time for other mathematicians to understand the content and use it for follow-up work.

On the 6th (local time), OpenAI compiled results across various fields, including algebra, geometry, and probability theory, and posted them on GitHub. On the 7th, the day after the release, three drafts with errors in their proofs were withdrawn, leaving 719, and the mathematics community is reviewing them.

Korean mathematicians who followed the release cited AI's ability to surpass the time limits of human researchers as a strength. Mathematics is highly specialized, making it hard for a single mathematician to broadly understand and use research from other areas. By contrast, AI can rapidly advance follow-up research by consolidating knowledge accumulated in different fields.

Seo In-seok, a professor in the Department of Mathematical Sciences at Seoul National University, said, "The limited time of human researchers has long been a practical bottleneck in mathematical research," adding, "This result shows AI has significantly cut the time needed for the next research stage by leveraging the body of work human mathematicians have built over many years."

He went on, "Just as it was hard for someone living in the Goryeo or Joseon eras to expect rapid national development within a lifetime, past mathematicians found it difficult to expect a quantum leap in their field within their own lives," adding, "With the advent of AI, the contribution a single mathematician can make will grow dramatically."

In practice, OpenAI's AI attempted to solve the problem of the "fluctuation theorem for the low-temperature Sherrington–Kirkpatrick model" based on existing human research. The problem concerns a model that mathematically represents a situation in which countless micro-magnets interact randomly, and requires determining how much a physical quantity called "free energy" changes at low temperatures.

According to Lee Ji-woon, a professor in the Department of Mathematical Sciences at KAIST who was studying the fluctuation theorem for the low-temperature Sherrington–Kirkpatrick model, the field long lacked even an approach to a solution. But in July this year, an idea first appeared that suggested the problem might be solvable, and OpenAI argued the problem could be resolved by developing that idea.

Lee said, "Rather than generating new ideas or research directions yet, AI is more of a tool that can efficiently and quickly use the ideas and diverse research results mathematicians have developed so far."

That does not mean AI's abilities stop at using prior work. Eom Sang-il, Chief Investigator (CI) of the Discrete Mathematics Group at the Institute for Basic Science (IBS), said the release also includes counterexamples showing long-standing mathematical conjectures are wrong, as well as new algorithms that were hard to imagine. A counterexample overturns the belief that a given mathematical claim always holds.

For example, OpenAI claimed it presented a case showing that while a three-dimensional space can be tiled perfectly without gaps using congruent shapes, it is impossible to arrange those shapes by repeating them under a fixed rule. This is a counterexample to the "periodic tiling conjecture," which holds that if space can be tiled without gaps by congruent shapes, then a regular, rule-based arrangement is also possible. It also presented a new algorithm that minimizes total completion time when distributing multiple jobs to three identical machines while preserving job order.

But there is room for improvement before mathematicians can understand and use AI's results. Lee said, "The writing is far too scattershot and there are many odd notations, so it will take a very long time for human researchers to understand and verify it."

Lee added, "Even reviewing a single paper takes considerable time, and releasing hundreds at once has increased the burden on the academic community to review them," emphasizing, "If AI is to contribute to mathematical progress going forward, it needs to improve how results are released."

Meanwhile, as AI's role grows, some said the mathematics community needs to rethink how researchers are evaluated. Seo said, "We have no choice but to consider how to evaluate the capabilities of human researchers in papers to which AI contributed," adding, "This could affect every evaluation system, from doctoral defenses to postdoctoral and faculty hiring, and tenure reviews."