Does AI-generated mathematics break the link between proof and understanding?

The essay argues that artificial mathematics, research papers produced with AI, pulls apart two things that have always come together in mathematics: the practice that produces understanding and the measurement that certifies it. Its central sentence is "Papers certify that a mathematician practiced the deepest form of our art and came to a complete understanding of a piece of mathematics." The worked case is Cheng, Liu and Gao's account of the Danus protocol, which proved a matroid-theory result and "wrote an equivalent proof" to Cheng's without access to his work. The essay reports that account as theirs and does not test it. Its argument concerns what the result is worth, not whether it is right.

The essay keeps two standards apart. A result is verified when someone confirms it is correct. It is understood when a person has built the argument, and the essay says that happens by writing: "If papers are produced without the human understanding that comes from writing them ourselves, they have little mathematical value to us, even if they prove 'important' propositions." It asks where understanding lives, offering "the work you read and referee, your napkin sketches, or the work that you yourself have written up," and answers that "AI tools cut humans out of our deepest mathematical experiences." The second half concerns journals. Knowledge already outruns absorption, and autonomously produced papers "will break the journal system." Tao is quoted proposing that prestige shift "to the humans who successfully verify and digest such proofs." The essay replies that checking "doesn't sound prestigious. It sounds boring." On this reading verification can be handed off, understanding cannot, and the field would not reward the handoff.

Against the nearest notes, the essay applies the decoupling argument of Does AI separate intellectual form from the thinking behind it? to a credential rather than a product. That note separates the outward form of intellectual work from the process behind it. The essay separates a paper, which is supposed to witness a process, from the process itself, so the damage lands on evaluation and not only on output. The essay also states the delegation claim that Is AI development already being handed to AI systems? makes from the lab side. It asserts that AI companies seek to automate AI research, but without the measurements Anthropic cites, so it is a critic's reading of intent rather than a finding. Its implied remedy, referees checking papers after they are written, is the gap that Can separating judgment from verification improve research paper reliability? addresses by building checks into generation. The essay does not consider that option.

The excerpt establishes a position, not a measurement. It gives no data on how often mathematicians use AI, how many papers are machine-written, or whether understanding actually drops when a proof is machine-generated. Its evidence is anecdotal: Cheng, Liu and Gao's experiment, Tsimerman's change of research direction, and an OpenAI claim about Fields Medals that the essay cites to an outside source. The essay's stronger program, "total opposition" to AI in mathematics with institutional organizing, follows from the understanding argument only if understanding cannot be recovered by other means, and the excerpt asserts that rather than showing it. At the strength the evidence allows, the verified-versus-understood distinction is worth keeping when a correct AI result arrives. The essay makes a case that a paper is weak evidence of understanding, but it does not say how understanding could be measured instead.