A misalignment of AI in mathematics

Who Should Read

A developer using LLMs who worries AI output may replace understanding and learning

Core Mechanics

• Math ability jumped in recent months
  • The declaration says LLMs can now solve major open problems in many fields.
  • It says AI companies' push to use math as a benchmark harms mathematics.
• Solving is a tool, not the goal
  • The real goal of math is conceptual understanding and insight.
  • Mass-producing true/false results could destroy fertile ground for new ideas.
• Rushed announcements skip the writeup
  • Solutions are announced without proper writeups, isolated methods, or citations of prior work.
  • This raises attribution and plagiarism questions, as in other creative professions.
• The human transmission chain may break
  • AI-conceived ideas only come alive if mathematicians develop them and fold them into the canon.
  • Without that, the human chain between mathematicians is lost.
• It's bigger than math

Community Reactions

• A mathematician is more optimistic
  • Mochizuki's abc proof was also huge and incomprehensible yet sparked conferences, papers, and discussions.
  • They suggest an AI Lean-verified proof would trigger similar community activity.
• Chess as a counter-precedent
  • Commenters recall 90s worries that computers would ruin chess, yet chess is now more popular and players are stronger.
  • They argue an answer oracle can help humans build more accurate insight.
• Coding experience shows creativity gaps
  • One engineer says Claude's 1000-line additions had no bugs 95% of the time but missed obvious performance, maintainability, and UX opportunities until prompted.
  • Another says AI routes around difficulty, so friction as a signal is lost to the human.
• The ask is unclear, and students struggle
  • Some say it's unclear what AI companies should concretely do differently.
  • Others report students now show less understanding and creativity because they skip the struggle.

How to Apply

• Add an understanding check to AI-generated code
  • Pick one AI-written PR today and summarize in three lines, in your own words, why it is built that way.
  • If you can't, don't merge it; ask the AI for its design rationale.
• Add an alternatives item to your review template
  • Require at least one human-written alternative on performance, maintainability, or scalability.
  • Commenters report AI often doesn't raise these opportunities on its own.
• Add no-AI exercises to junior onboarding
  • Schedule a weekly session of debugging without AI.
  • This applies the declaration's point that struggling with problems builds understanding.

Try It

# Paste into Claude or ChatGPT (a simplified version of the declaration's understanding-vs-results idea, applied to code review)

You are a senior reviewer. The code below was written by AI.

{{code diff}}

Answer the following.
1. Summarize the problem this code solves in 3 lines
2. Give 3 alternatives this design may have missed (performance, maintainability, scalability)
3. Write 5 questions I can use to check I truly understand it (do not give the answers)

Next step: answer the 5 questions yourself and ask for grading.

Terminology

Related Papers

• The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge

  Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.

• 90% of US working hours are spent on tasks not covered well by AI benchmarks

  Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage

• Postgres 19 delayed after AI reviewers find more bugs than humans can fix

  LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted

• Show HN: Gemini 4 Argon plays StarCraft

  LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.

• A Plausible Explanation Is Enough to Mislead Multimodal Judges

  Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.

• Neural networks are clean

  Looking at neurons directly shows most have one main job, challenging superposition.

Related Resources

• https://mathandai.org/
• https://news.ycombinator.com/item?id=49662116
• https://fee.org/articles/the-candlemakers-petition/
• https://encyclopediaofmath.org/wiki/Finitism