과학/건강

OpenAI's announcement on October 26, 2026, was the clearest manifestation of this change. The company stated that its AI model had produced proofs for decades-old difficult problems in mathematics (Local Smoothing Conjecture), computer science (Exponential Time Hypothesis, ETH), and physics (existence proof of interacting quantum field theory, related to Yang-Mills and QCD).

According to reports, the physics section is linked to some of the Millennium Problems, and a group of experts, including 2004 Nobel laureate in physics David Gross, is leading the verification. The mathematics and computer science sections have been formalized and released in Lean, while the physics section is being rewritten by experts due to its complexity. Interpreting this announcement as "AI has surpassed human limitations" is a misunderstanding of the conditions for the establishment of scientific knowledge. Knowledge is not merely the presentation of a correct answer; it is established only when logical consistency, accurate correspondence with the original problem, consistency with existing research trends, and the interpretation of the meaning of the results within a broader theoretical framework are all satisfied.

This becomes clear when we look at the three achievements in detail. The local smoothing conjecture is about the mathematical property that waves do not concentrate excessively at certain points. If proven, it would sequentially solve the Kakeya conjecture and Stein's restriction conjecture. Although it is claimed that AI introduced auxiliary partial differential equations to provide a proof in all dimensions, human mathematicians must reconstruct and explain how its core idea connects with existing research trends and the extent of its new insights. ETH's refutation overturns a hypothesis about the time complexity of a specific computational problem, and it is significant that Russell Impagliazzo and Ramamohan Paturi, who originally proposed the hypothesis, were included in the verification group. Human interpretation and further research are needed to understand the practical implications of the algorithm discovered by AI and its relationship with the P vs NP problem. In the physics section, the mathematical existence proof of interacting quantum field theories, including QCD which describes the interaction of quarks and gluons, is directly related to the foundation of theoretical physics. However, the fact that the data has not yet been fully disclosed and the expert group has taken the stance of rewriting it themselves suggests an unwillingness to accept the results produced by AI as knowledge as-is. The precedent of errors and verification burdens being pointed out in AI-authored mathematical papers massively released by OpenAI in the past also supports this caution.

This pattern is in line with the evolution of tools shown in the history of science. When the telescope expanded the scope of observation, astronomers manually integrated observational results into theories. Even when computers took over numerical calculations, the interpretation of physical meaning and the judgment of model validity remained with humans. Even when AlphaFold achieved a breakthrough in protein structure prediction and received the Nobel Prize, the process of experimentally verifying prediction results and assigning biological meaning was the domain of human researchers. AI is a qualitative leap in that it can generate hypotheses and perform formal proofs much more actively than previous tools, but that leap is closer to 'automation of exploration' than 'automation of understanding.' Formal verification tools like Lean are excellent at filtering out mechanical errors, but the ability to judge “whether this proof properly captured the essence of the original problem,” “how this result connects with other unsolved problems,” and “what new questions this knowledge opens up” still lies with humans. AI's achievements operate on the accumulation of human knowledge—training data, problem settings, formal verification tools, and networks of experts participating in verification.

The symbiosis between humans and AI in the future is therefore likely to evolve beyond a simple division of labor, where 'AI solves and humans verify,' into a more sophisticated, mutually reinforcing relationship. AI will propose candidates that humans would find difficult to reach in a short time within a vast search space, and quickly check for formal consistency. Humans will set the meaning and priorities of problems, critically reconstruct the results produced by AI, and connect them to the real world through experimentation, observation, and theoretical integration. Furthermore, it will remain the human's role to clarify the limitations and preconditions of AI outputs, and to reflect on the social, technological, and ethical implications of those results. OpenAI's approach of not pursuing academic credit and including original hypothesis proposers or authoritative researchers in the verification process can be seen as an attempt to experiment with the initial norms of this symbiosis. If this model takes root, the speed of scientific research will accelerate, but the reliability and transmissibility of knowledge will depend on the quality of the human-led processes of verification, reconstruction, and interpretation.

For this symbiosis to be sustainable, three conditions are crucial.

First, the systematic expansion of verification infrastructure. If the quantity of candidates generated by AI increases exponentially, the verification capacity of human experts becomes a bottleneck. Without the standardization of public repositories, the advancement of formal verification tools, and academic recognition and financial support for verification work, 'verification overload' could actually slow down the pace of science.

Second, maintaining interpretability. If black-box results accumulate, the scientific community risks relying on 'correct results without knowing why they are correct,' which could lead to a long-term reduction in theoretical intuition and the ability to ask creative questions. The process of humans rewriting and explaining AI results is not merely confirmation, but an essential step for the transmission of knowledge and the growth of the next generation of researchers.

Third, the institutionalization of trust and responsibility. Bias or selective disclosure can occur when AI companies choose problems and produce results. The autonomy of independent verification groups must be guaranteed, data must be sufficiently disclosed, accountability for results must be clarified, and agreed-upon norms regarding the scope and timing of disclosure must be established.

Science has always evolved with the development of tools. AI is entering the next stage of these tools, but no matter how powerful the tools become, the judgment and integration that transform what the tools produce into 'meaningful knowledge,' and the reflection on the implications of that knowledge for nature and society, still remain within the human domain. What this announcement suggests is that an era has arrived where AI is tackling decades-old challenges, and at the same time, that scientific progress only occurs when that 'hand' is combined with human eyes, minds, and a critical community.

Symbiosis has already begun, and its quality depends on how diligently we design systems for verification, understanding, and responsibility. The moment when AI surpasses human limitations has not yet arrived. Rather, a new stage of science is simply opening up where humans and AI complement each other's limitations and together push those boundaries. In that stage, what humans must not lose is not speed, but the ability to ask why those results are important and what they mean for our world.