Who Still Checks Expert Work
A clean proof, report, or benchmark result can now arrive with the tone of finished expert work before the slower part begins: inspection. That is the concrete risk in front of teams this week. If expert-level AI work becomes easier to produce than to evaluate, the first thing at risk is not only correctness. It is the human deliberation that makes expert judgment worth trusting.
A widely circulated profile said Terence Tao now warns that AI could destroy mathematics' deliberative process, as Techmeme posted in April 2025. The point is not just that a model may help write a proof. The deeper problem is that fewer people may still do the arguing, checking, and comparison that turns a proof into something a field can rely on.
The same pressure appears in how expert benchmarks are discussed. A same-day post said benchmarks of expert-level work are hard to pull off and that major benchmarks such as ARC-AGI, GDPval, and METR Long Tasks are saturated, as Ethan Mollick posted in April 2025. Saturation matters because it weakens a basic control. When the benchmark no longer separates systems clearly, polished performance can look settled before the evaluation method itself is settled.
A second same-day post about a successor to GDPval made the evaluation problem more concrete. It said the successor still needs human baselines, that GPT-6 Pro spotted some grading errors, and that 0% scores can hide economically meaningful differences in success rates, as Ethan Mollick posted in April 2025. That is not a small technical footnote. It suggests the inspection layer is under strain at the same time the generation layer gets smoother.
This is where accountability can slip. If the output looks expert and the benchmark looks exhausted, people can inherit a result without inheriting the checking that gave expert work its authority in the first place. In mathematics, that threatens deliberation. In ordinary office and technical work, it threatens review, signoff, and the record of who decided what.
The practical response is plain. Keep a human review step explicit instead of implied. A decision ledger is a short written record of who checked an AI-assisted result, what baseline they compared it against, and what they accepted or sent back. At work, it looks like a named reviewer, a named comparison point, and one written note about the check that was actually performed.
That matters because the control here is not to ban AI output or pretend every result needs full manual reconstruction. The control is to make the evaluation visible again. If a proof, report, or benchmark result will guide a meeting or a decision, name the human baseline before the polished answer gets treated as settled. If there is no baseline yet, say that plainly too.
A useful baseline can be simple. It might be a prior human-written report, a known manual grading process, or a domain expert's written standard for what counts as complete. The point is not that a baseline makes the answer true. The point is that it restores a comparison a person can inspect and defend.
This week, require one written check before expert-sounding AI work is treated as finished, and compare it against a named human baseline when one exists.
Human Runtime Core™ is designed for the mental edge you need to survive in an AI-run world.