A Thousand Mathematicians Sign Pact Not to Use AI; 654 May Have Become the Cleanest Control Group

More than 1,000 mathematicians have joined an organization called the Association for Human Mathematics. The public commitments are very specific: not to provide technical labor, mathematical knowledge, consulting, or promotion for commercial AI companies; not to publish AI-generated mathematical text, including papers, referee reports, lecture notes, textbooks, and books; and 654 people have further joined a no-AI group, using no AI models at all in their research—not for finding proofs, exploring examples, or posing research questions.

The most notable number in this set is 654. Someone in the replies said this is the cleanest control group in mathematics and should be pre-registered. It's a bit mean, but the direction is right. Between mathematics done with AI and mathematics done without AI, who will understand more deeply and whose questions will be more interesting a few years from now? These 654 people happen to provide an observation window.

In the same discussion thread, Hesamation uses chess as a comparison. He found a passage from Magnus Carlsen. Carlsen says neural networks have greatly improved our understanding of chess. He is very humble about engines: the problem is that my understanding is not good enough, and this is very useful. He also says he rarely plays against engines, because they make him feel stupid and useless.

Put these two passages together, and the conflict is obvious. Mathematicians worry that AI can help find proofs without increasing human understanding of mathematics. Carlsen does not deny that engines make him look stupid, but he acknowledges that engines have improved understanding. Where is the difference? One reply says that a chess engine gives an evaluation for every position, so you can see where the judgment is off; a proof returns right or wrong, just one bit. Carlsen learns from numbers and moves, not merely by accepting a move.

Another reply adds the time cost: players do not get understanding for free. They stare at engine lines for many years, and only then do strange moves slowly become understandable. If mathematics follows the same path, it too will first serve as a proof checker, then become a source of intuition. The risk is that the commitment shuts down the second stage as well.

Chess experience cannot be transferred directly. Some say chess is a closed 64-square board, while intermediate processes in mathematics open new paths; others say mathematicians are not performers, and AI replacement is different from that of chess players. These rebuttals hold. But Carlsen's humility is still worth looking at. He does not play against engines because it is uncomfortable; he studies engines because they can show him where his intuition is wrong. This distinction is more useful than taking sides.

There are also several typical voices in the replies: carriage drivers rejecting cars; the 654 will be crushed by graduate students using Codex; it is 100 percent ego that slows the field down; handmade mathematics is like an Etsy label; and some think this is virtue signaling. Sharp, but it points to a question: the group that rejects AI may precisely provide the observation window. Between mathematics done with AI and mathematics done without AI, who will understand more deeply and whose questions will be more interesting a few years from now? The data will speak.

The facts that can now be confirmed are few: AHM exists, more than 1,000 people have joined, and 654 have committed not to use AI in their research; Carlsen said those things. All remaining inferences should slow down. What pain point AI solves first in mathematics—finding proofs, checking proofs, giving examples, or posing new questions—different answers will lead to different conclusions. Treating AI as an answer generator and treating AI as a feedback-dense sparring partner are two different things. Chess has already played this out once: first it was used to verify lines, later to learn intuition. Whether mathematics will repeat this is worth watching with those 654 people in mind.

The original discussion is on X: Hesamation's tweet, and the relevant AHM citation is here.