The evidentiary problem of AI-generated content is not a disclosure problem. It is an epistemological one. Every major governance framework — the EU AI Act's Article 50, proposed FRE Rule 707, the Take It Down Act, platform watermarking policies — addresses synthetic content governance as a labelling problem: if generated content is properly disclosed, tagged, or marked, downstream harms can be managed. This Article argues that framing is structurally mistaken. The relevant distinction between observed and generated content is not a property that can be attached to content after production; it is a classification of epistemic status — specifically, the presence or absence of the constitutive causal connection to observable reality that makes content capable of anchoring knowledge claims. No label, watermark, or disclosure requirement can restore what is structurally absent. This Article traces the history of evidentiary authentication through three transitions — physical artefact, photograph, digital file — demonstrating that each produced an authentication gap that law addressed procedurally rather than formally. The current transition to AI-generated content represents a structural rupture, not merely a further step in the sequence: synthetic content lacks not a causal chain that can be documented but the causal chain that makes evidence epistemically significant. Five categories of evidentiary failure produced by this rupture are identified and analysed: identity attribution without observational grounding (the Taylor Swift deepfakes), authority attribution without testimonial grounding (the Biden robocall), harm attribution without causal grounding (the AI-CSAM crisis), authentication without classification infrastructure (the deepfake defense), and authorship attribution without creative grounding (the AI authorship problem). The Modulign Standard v3.0 — a published open specification for Dimensional Address Grammar for Observable Reality (DAG-OR) — provides the formal infrastructure that closes the classification deficit. Its VR/·:SYN classification primitive is the first formal epistemic primitive for synthetic content, encoding the causal deficit at the classification level rather than as metadata. Its Address-Theoretic Non-Accidentality Principle (ATNA) provides the first structural dissolution of the Gettier problem as applied to observational evidence, guaranteeing that a properly classified observation cannot be accidentally true. Its confidence threshold architecture maps emergently onto legal standards of proof. Its ^EVID append-only ledger constitutes the chain of custody from the moment of classification. A clause-by-clause analysis of Article 50 identifies five structural gaps that DAG-OR closes without displacing the regulation. Proposed FRE Rule 901(b)(11) amendment language is provided. A unified four-step implementation framework concludes the Article.
Gonzalez, V. (2026). THE CLASSIFICATION DEFICIT Article 50 of the EU AI Act, the Epistemic Gap It Cannot Close, and the Formal Standard That Does. Zenodo. https://doi.org/10.5281/zenodo.19578570