The rapid advancement of large language models (LLMs) has introduced a new era for mathematical research. These systems have demonstrated the ability to generate complex proofs, such as those involving the Navier–Stokes equations and the quasi-Riemann Hypothesis. However, as these tools become more integrated into the scientific ecosystem, they raise fundamental questions about the nature of human understanding and the future of mathematical progress. The core tension lies between a 'golden age,' where AI empowers a new generation of scientists to explore previously unreachable frontiers, and a 'dark age,' where scientific output increases but becomes increasingly concentrated in a way that is alien to human cognition.
One significant technical hurdle in this transition is the 'translation' problem. While LLMs can generate correct proofs in formal languages like Lean—a proof assistant that allows computers to verify arguments—the natural language translations of these proofs often become increasingly messy and less faithful as the complexity of the proof grows. This creates a scenario where a machine might produce a correct formal proof that is too long or complex for a human to directly read or interpret. This limitation highlights a crucial distinction in the field: mathematics is not merely a collection of solved problems, but a human cultural artifact rooted in how humans cognize space, time, and quantity. Without human understanding, the value of a proof may be reduced to a mere benchmark for AI performance rather than a contribution to human knowledge.
Beyond mere proof generation, the most significant opportunity for AI may lie in 'mathematization'—the act of inventing new mathematical languages and conceptual frameworks to make reality intelligible. Historically, breakthroughs by figures like Newton, Riemann, and Grothendieck were defined by their ability to create these new frameworks. While it remains uncertain if LLMs can perform this level of conceptual invention, the ability to automate the 'commodity' of proof-checking could free humans to focus on these higher-level creative tasks. However, this shift also necessitates a re-evaluation of authorship and peer review. As AI-generated work draws on vast bodies of human literature, the mathematical community must develop new systems to ensure that credit is attributed correctly and that the shared mathematical culture remains accessible and meaningful to human researchers.