Type: article
Abstract: Agentic coding systems now produce software faster than the workers accountable for it can read, evaluate, or trace. We theorize this shift through epistemic debt: the latent stock of work artifacts whose verification has been deferred by an accountable worker in a human-AI collaboration. We formulate the construct for AI-mediated knowledge work, distinguish three strategies — accept, negate, triage — for managing it, and examine it through 40 controlled Claude Code sessions, a repair experiment on 32 programs, a thematic analysis of 4,570 Reddit items, and six recorded sessions. We found that a community-derived constraint prompt produced a small, non-significant reduction in drift. The repair experiment showed when debt becomes consequential: repair burden rose when a fix required relational knowledge the code could not supply. The corpus reveals a 15-practice repertoire dominated by selective verification (92% triage). Tools and institutions designed for verification, not only production, are the next frontier.
Authors: Zhang, Shiyan, Nickerson, Jeffrey V
Institutions: Lancaster University, Stevens Institute of Technology
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222909312