认识论债务：AI 依赖型软件工程的概念框架

View PDF

Abstract:This paper proposes a conceptual framework for Epistemological Debt, a liability that AI-dependent software engineering creates and conventional engineering metrics miss: the share of a system's behavior that no maintainer can currently explain. The debt accrues in two ways concurrently inside a single workforce. Maintainers with no theory of a system ship generated changes without forming one, and maintainers with a theory stop deriving changes themselves, so the system drifts away from a theory that is still intact. Neither is the attentional lapse the automation complacency literature describes, and the paper explains how to tell them apart. Against technical debt, tacit knowledge, and Naur's theory-building account, the contribution is this: both mechanisms share a cause, and running concurrently, they close the transfer window that earlier accounts of deskilling under automation relied on. The framework is stated as six propositions and two open hypotheses, with an evaluation design that separates the mechanism from the alternatives. Amazon's 2025 and 2026 service incidents illustrate the governance gap the framework predicts; that reading is labeled as interpretation. Proposed interventions are sorted into those that build comprehension and those that only limit risk; none is claimed as validated.

Comments:	10 pages, 2 tables, 1 figure. Submitted to IEEE Software
Subjects:	Software Engineering (cs.SE); Computers and Society (cs.CY)
ACM classes:	K.4.1; D.2.0; I.2.0
Cite as:	arXiv:2604.26855 [cs.SE]
	(or arXiv:2604.26855v4 [cs.SE] for this version)
	https://doi.org/10.48550/arXiv.2604.26855 arXiv-issued DOI via DataCite

Submission history

From: Frank Ginac [view email]
[v1] Wed, 29 Apr 2026 16:20:25 UTC (132 KB)
[v2] Sun, 3 May 2026 21:34:35 UTC (288 KB)
[v3] Tue, 18 Aug 2026 01:38:21 UTC (370 KB)
[v4] Thu, 15 Oct 2026 19:23:34 UTC (243 KB)