Year: 2025
Type: article
Abstract: In recent years, generative models (LLMs) have become an "accelerator" of office work and software development: texts, reports, minutes, analytics, code, and documentation are produced and reworked faster than ever. However, this acceleration has been accompanied by the accumulation of a new class of debt — verification debt — which manifests not as immediate system failure but as organisations gradually losing the ability to verify, explain, and reproduce their decisions. We propose the conceptual framework of "knowledge rot" and "shadow metrics" for diagnosing this debt, linking the phenomenon to (i) recursive reworking of artefacts ("generation → summary → summary of summary"), (ii) the proliferation of opaque code and documents, (iii) consensus hallucinations in deliberative processes, and (iv) drift in models, data, and prompts. The article contributes: (1) a taxonomy of mechanisms by which corporate memory degrades; (2) a formalisation of a "shadow quality circuit" by analogy with measuring the shadow economy, including a MIMIC approach to estimating the proportion of unverified AI artefacts; (3) the design of an internal "AO module" (Algorithmic Oversight module) as an independent decision-audit circuit; (4) a set of operationalisable metrics (Human Verification Rate, Data Provenance Score, Algorithmic Drift Index, Recursive Depth Index, and others) together with implementation protocols.
Source: AI DECISIONS
Author: Yury Korolev
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7128360864