Studying proof data with proof theory: the case of AI-generated mathematics

Overview & Synopsis

I present a method for analyzing large collections of formally verified mathematical proofs (e.g. OpenAI's) grounded in some historical and philosophical musings on mathematics in the age of AI. Many of the quantities involved, such as the size of a proof term or the number of redundant steps, are already computed by work on improving automated provers, where they serve as targets. My aim is to turn them into measurements of proofs themselves: compared against a reference library and against each proof's own cut-free form, checked against theorems they must obey, and controlled for how the proofs were compiled. With these instruments, I measure how (much) machine-generated mathematics diverges from human mathematics, and find that AI-written formal proofs realize far less of the logical compression available to them than human mathematicians do (31% against 73% in mathlib, and 55% against 86% on proofs that call no automation). These figures are broadly in line with reports from those who build provers and proof-improvement tools that AI-written proofs are often longer than they need to be.