Who signs off AI proofs?
The unsettling part is not that AI can produce expert-looking work. It is that someone still has to decide when it counts.
This month, the volume and pace of claims made that plain. The Rundown AI posted in April 2025 that Axiom completed a fully machine-checked version of the Classification of Finite Simple Groups in 24 days with a team steering an AI agent swarm. The Rundown AI posted in April 2025 about multiple AI math and science results arriving in quick succession, including a claim that Claude disproved the Exponential Time Hypothesis using 14M output tokens and no human input. Simon Willison wrote in April 2025 that October looked like another inflection point for coding agents and that Claude Opus 5.5 could build video games that are actually fun.
Taken together, those posts describe a real shift in the work around technical output. Drafting is getting cheaper and faster. The bottleneck moves somewhere else. In math, software, and research, the scarce step is no longer producing a plausible artifact. The scarce step is deciding what review is sufficient before a team, a lab, a product group, or a manager marks that artifact as accepted.
That is why the risk is not only false claims. It is also unowned acceptance. When machine-generated proofs, code, and research notes arrive faster than calm review, authorship and accountability can separate. The work appears before the standard for accepting it does. Ethan Mollick wrote in April 2025 that AI forces academics to rethink how science is done because AI does not share the same paper-production constraints humans do. skdh wrote in April 2025 about the tension around AI-written proofs by contrasting demands for human checking with pressure to release proofs immediately.
A decision ledger is a written record of what must be checked, who checks it, and what blocks approval; at work it looks like a short acceptance note attached to a proof, pull request, or research claim before anyone marks it as done.
The durable human job, then, is not to race the model at first-draft speed. It is to set the gate. A careful team can still decide what counts as enough verification for the kind of output in front of them. That decision changes with the work. A coding agent patch may need tests, reproducibility, and a named reviewer. A proof may need machine checking plus human review of assumptions and statement scope. A research claim may need source inspection, replication steps, and explicit sign-off on what remains uncertain.
This is not a call to slow everything down. It is a call to make acceptance criteria visible before the output lands. If AI can produce ten candidate arguments, functions, or claims before lunch, then the team needs a prewritten answer to a narrower question: what must be true before one of these enters the record. The person who sets that answer still matters because the acceptance standard is now a primary form of expertise.
Lex Fridman wrote in April 2025 that hard problems in mathematics, computer science, physics, and science were being solved by AI over a period of weeks, calling it either a crisis, a triumph, or both depending on perspective. The practical reading is simpler. When output expands, review cannot remain implicit. If nobody defines the check, the gate defaults to speed, attention, or optimism. None of those is a sign-off standard.
Set the check this week
Before you use any AI-generated result this week, define one explicit acceptance check. Write down three things:
- what must be verified
- who must verify it
- what would block approval
Keep it specific to the artifact in front of you. For code, that may be tests, security review, and rollback conditions. For a proof, that may be machine verification, assumption review, and a named human approver. For a research claim, that may be citation checks, reproduction steps, and a statement of what remains unverified.
Human Runtime Core™ is designed for the mental edge you need to survive in an AI-run world.