AI systems increasingly perform the local work of drafting, coding, summarizing, searching, planning, and revising. This working paper argues that such systems do not eliminate human expertise. They change where expertise appears.
The paper introduces operator judgment as the situated human capacity to specify the object of work, target the relevant verification, intervene when the system errs, and govern whether outputs may advance into downstream use. It develops this framework using Anthropic’s Claude Code report, Agentic Coding and Persistent Returns to Expertise, as an empirical hook. The report treats expertise as task-specific rather than occupational and classifies it through setup specificity, verification type, and direction of correction. This paper interprets those signals as observable components of operator judgment.
The paper situates operator judgment within older traditions of expertise studies, naturalistic decision-making, situation awareness, supervisory control, human-AI interaction, distributed cognition, professional judgment, and verification and validation. It also uses transcript-sufficient research systems as a case where state governance becomes visible: AI-mediated outputs may become evidence, working state, proposals, review objects, canonical references, or public artifacts.
The central claim is that AI makes execution easier, but judgment more important. Expert AI use should not be reduced to prompt skill, coding skill, or job title. As AI systems produce more candidate work, the scarce human capacity is knowing what the work is, what would make it count, when to stop or correct the system, and what the result is allowed to become.
This is Article OJ-01 in the standalone Operator Judgment series: Expertise, Verification, and Governance in AI-Mediated Work.
Publication Date: 2026-06-18