Expert performance on isolated occupational tasks is insufficient to establish expert performance of the occupation.
Assessed atomic
The budget was tight, so the claim was kept atomic: no subclaims and no named arguments. It was assessed supported (credence 0.85) on construct-validity reasoning and the gap between GDPval and the Remote Labor Index. No sources were recorded, and a later pass should add instances.
Assessed Supported
verdict confidence 0.75 · credence 0.85
The claim holds that doing isolated occupational tasks at expert level does not, on its own, show that someone, or some system, can perform the occupation at expert level. It is mainly invoked in debates over AI capability benchmarks. The reasoning behind it is strong. An occupation is more than a list of separable deliverables. It also involves choosing and scoping problems that arrive ill-specified, coordinating with other people, carrying context across weeks, taking responsibility for outcomes, and recovering from errors. A benchmark made of discrete, well-specified tasks with clear grading samples only part of that. The empirical record points the same way. Task-level evaluations such as GDPval report AI output near expert quality on sampled deliverables. End-to-end evaluations of whole remote-work projects, such as the Remote Labor Index, report much lower success rates. That gap is what one would expect if success on tasks does not transfer to the job. The same logic is familiar in human credentialing: passing a licensing exam is treated as necessary but not sufficient, and supervised practice is required before independent practice. The disagreement is mostly about degree. Defenders of task-based measurement can reply that broad, representative task coverage, together with evidence that the parts combine, would come close to establishing occupational performance. The claim's word "isolated" excludes that case. Read as stated, the claim is well supported.
Claim entered the graph