From Deepfakes to Machine-Mediated Reality: Epistemic Decoupling and the Duty of Evidential Grounding in UK Public Law

Abstract

. This paper argues that UK public law contains no sufficiently developed principle governing decisions based on machine‑mediated states of affairs: situations in which an automated system does not merely observe or classify the factual basis of a decision, but materially alters that factual basis before supplying the evidence of it. As AI systems increasingly act in the physical world—inspecting, adjusting, and maintaining the infrastructure that later decisions depend on—the system that changes a state of affairs often becomes the primary source of evidence about it, closing an evidential loop that existing doctrine was not built to open. The paper distinguishes three objects of machine‑mediated governance concern: machine‑generated representations, machine‑generated actions, and machine‑generated or machine‑modified states of affairs. UK public law addresses the first two through evidential and product‑safety doctrine, but not the third. Existing doctrines of procedural fairness, rationality, disclosure, and evidential reliability—including the presumption that a computer system was operating correctly, and the duty of candour in judicial review—provide partial safeguards, but they do not consistently require preservation of a baseline state, an intervention history, or an independent means of verification. Drawing on computational‑complexity analyses of transformer‑based language models, the paper argues that the missing independent check cannot safely be assumed to come from another system of the same kind, and must instead be secured through preserved records and institutional safeguards. Using the Post Office Horizon litigation and recent judicial commentary on automated decision‑making, it proposes a duty of evidential grounding: a requirement that public authorities take reasonable and proportionate steps to preserve the records needed to reconstruct the factual basis of a materially consequential decision. It develops a graduated machine‑mediated grounding test, proposes an adverse‑inference consequence for unjustified failures of preservation, and offers a model statutory clause as a possible codification. • Expanded Section 3 with new discussion of Floridi’s “decoupling of agency from intelligence,” clarifying how this differs from the evidential decoupling addressed in the paper. • Added new footnote (Floridi 2023; Lozano Ortega 2025) explaining conceptual distinctions between types of decoupling. • Strengthened Section 4 with additional explanation of environmental “envelopes” shaped around machine operation, drawing on Floridi’s account of automation success. • Updated Section 5.1 to include Vaswani et al. (2017) as the foundational citation for transformer self‑attention complexity. • Clarified Sikka & Sikka’s 2025 argument by referencing the third version of their preprint (15 July 2025) and adding detail on composite systems and multi‑model capability. • Added new empirical evidence in Section 5.3 from Chandra et al. (2025) on deepfake detector degradation in real‑world conditions. • Strengthened Section 5.6 with clearer articulation of the independence requirement: a second model is not an independent evidential route unless it accesses facts, not just outputs. • Improved transitions and tightened language across Sections 3–6 for clarity and precision. • Minor structural edits to align terminology (“machine‑mediated states of affairs,” “preserved baseline,” “intervention log”) consistently throughout the text. • Updated citations and footnotes to reflect new sources and corrected publication years.