[Commentary] Mathematics in the Age of AI: From Human Intuition to "Drive-By Proving"

TL;DR

The boundary between human mathematical genius and machine computation is blurring. In this provocative essay, Jeremy Avigad (Director of ICARM) details how AI has moved from "laughable" errors to solving research-level theorems in under four years. The transition from the International Mathematical Olympiad (IMO) to real-world conjecture solving signals a paradigm shift: we are no longer just using computers to calculate; we are using them to think.

Perspective: The End of "Hiding Places"

For decades, mathematicians felt safe behind the "moat" of abstraction and intuition. While AI conquered Chess and Go, the rigorous, multi-step logical deduction required for a 50-page proof seemed uniquely human.

Avigad argues that these "hiding places" are vanishing. Recent developments at the Institute for Computer-Aided Reasoning in Mathematics (ICARM) show that AI agents are now:

• Closing formal proofs: Completing complex formalizations (like the E8 lattice optimality) that previously required months of human labor.
• Creative informal reasoning: Solving unpublished research problems with "beautiful" and "completely correct" arguments.

The "Drive-By Proving" Phenomenon

One of the most striking insights in the essay is the cautionary tale of the E8 lattice formalization. A team of human mathematicians spent years building the "scaffolding" (a blueprint for a proof). Suddenly, a corporate-backed AI agent ("Gauss") swooped in, utilized that human-built infrastructure, and "closed" the proof in secret—a tactical move Avigad calls "drive-by proving."

(Note: This conceptual diagram represents the workflow where humans provide the high-level strategy (blueprints) and AI executes the low-level logical tactics.)

The Risk: If AI "finishes" a project, does the human incentive to understand the why disappear? If an AI produces 10,000 lines of "spaghetti" code that verifies a theorem, have we actually gained mathematical knowledge, or just a green checkmark?

Methodology: The Neural-Symbolic Synthesis

The success observed isn't just from "Larger LLMs." It comes from the interplay of three distinct technologies:

1. Formalization: Proof assistants like Lean and the Mathlib library.
2. Symbolic AI: SAT solvers and automated reasoning that handle combinatorial explosions.
3. Neural Networks: LLMs that act as "navigators," suggesting the next move in a proof based on vast libraries of existing literature.

Why does this work? Because mathematics provides a strong signal for correctness. Unlike creative writing, a mathematical proof can be compiled and checked. This feedback loop allows AI to self-correct, eventually outstripping human speed in technical execution.

Experimental Reality: AI vs. Research Problems

In a landmark experiment conducted in early 2026, eleven mathematicians set ten challenge problems for AI.

• The Baseline: Standard models like GPT-4 were historically poor at this.
• The SOTA: Google DeepMind's Aletheia solved 60% of them.

(Graph indicating the rapid ascent of AI performance from 2022 to 2026, saturating student benchmarks and moving into research frontiers.)

Critical Analysis: Is Mathematics Obsolete?

Avigad rejects pure doomsaying. Instead, he proposes that mathematics is the solution to the "black box" problem of AI.

• Auditable Intelligence: Mathematics is one of the few fields where we can demand a step-by-step artifact of justification.
• The Architect Role: Mathematicians must stop being "calculators" (who compete with machines) and start being "architects" (who define the problems and verify the AI's solutions).

Limitations & Challenges

The essay identifies a significant threat to mathematical pedagogy. If students can solve any homework problem with an AI agent, the traditional "service course" model for engineering and business may collapse. We need a new way to teach "core intuition" in an era where the "tedium" of calculation is dead.

Conclusion: Show Up and Get to Work

The future of mathematics isn't "Human vs. Machine"; it's a "Human-Machine Hybrid" where the human provides the values, the questions, and the ultimate oversight. If the community ignores AI, they risk becoming obsolete. If they embrace it, they may finally solve the Langlands program or P vs NP.

The message is clear: Mathematics will thrive only if mathematicians "own" the technology rather than fearing it.