Work

Mathematical AI and Epistemology

Year: 2026
Type: preprint
Abstract: In 2025, both Google DeepMind and OpenAI reported that their large language model (LLM)-based AI systems had achieved gold medal level performance in the International Mathematical Olympiad. In addition, together with mathematicians, DeepMind and others have made progress in solving research-level mathematical problems with their AI system. These developments suggest that in the near future important parts of mathematical theorem proving may be outsourced to LLM-based AI systems, including what AI developers call large reasoning models (LRM). However, the LRMs still commonly make errors in simple reasoning tasks, prompting questions about their fit as mathematical (and other) tools. In this paper, I take first steps toward establishing an epistemology of artificial mathematical reasoning, in order to set up a systematic framework for studying mathematical reasoning in AI systems. I divide the topic into two questions. First, in the “knowledge of” question we ask how we can identify reasoning ability (if any) in AI systems. I argue for the notion of robustness as an identifiable characteristic. Second, in the “knowledge through” question we ask how mathematical knowledge generation takes place with the help of AI reasoning systems (if any). I present two scenarios, the interactive theorem prover (ITP) and autonomous automated theorem prover (AATP), and analyse mathematical theorem proving in them. Finally, I will consider the consequences such AI mathematical systems would have for epistemic agency of human mathematicians. How will the role of mathematicians change if an important part of theorem proving can be outsourced to AI systems? Could an AI system itself be considered a mathematical epistemic agent?
Language: English
Open Access status: green
OpenAlex ID: W7221271795
Cites: 50
Cited by:
Related to: 10