When the Source Disappears: Epistemic Supply-Chain Security and the Cognitive-Security Stakes of AI-Mediated Knowledge
Generative artificial intelligence increasingly sits between people and the sources from which they build an understanding of the world. It summarises, retrieves and synthesises, so that a representation is often encountered before, or instead of, the evidence behind it. Research has examined misinformation, hallucination, synthetic media and model collapse. This paper addresses a different problem: what happens to the socially accessible record when AI mediation grows while the provenance, accessibility and independent verifiability of the underlying human sources decline. The paper introduces epistemic supply-chain security (ESCS): the protection of provenance, integrity, traceability, plurality and verifiability along the path from primary human sources through digitisation and generative AI to the knowledge from which beliefs are formed. It develops the fragmented-truth problem, in which authentic pieces of information produce a misleading whole through selection, omission, decontextualisation, weighting or amplification, and recursive epistemic degradation, in which repeated synthetic mediation increases the distance between accessible knowledge and its authenticated origins. The argument is that this distance is a security vulnerability. Where independently accessible sources decay, the capacity to return to them and check the record decays with them. That capacity is what allows falsehood, omission and selective framing to be corrected. Its loss does not require a hostile actor, but it may lower the cost of one: an adversary need not insert falsehood if it can shape the environment through which a population determines what is true. Preservation, provenance and source accessibility are therefore matters of cognitive security, not only of archival practice. The paper further argues that verifiability has a behavioural condition at the point of reception. A reader who has lost track of where a claim came from can still recover it through citations, logs, search or external records, but may be less likely to attempt the recovery, and source-monitoring research shows that tracking is unreliable and susceptible to AI-mediated interaction. The paper states four falsifiable propositions and specifies the studies that would test them.
Authors
- Sarah Jane Gardner
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24222084
- Primary Topic
- Misinformation and Its Impacts
- Type
- preprint