Article 50 makes the export pipeline carry the AI provenance

The mark and the notice are different artifacts, and they fail in different places. A notice gets rendered at publish time, and each channel renders it its own way. The mark has to be in the asset. A dev.to explainer of the rules puts the awkward case plainly: a team may need to preserve the information needed for machine-readable identification when it exports a video for several platforms ^[11].

Coverage differs too. Machine-readable identifiability attaches to AI-generated or manipulated material, while the user-facing disclosure triggers where the content is a deepfake, meaning AI-generated or manipulated image, audio or video resembling real people or events, or where covered text concerns matters of public interest ^[4][2][3]. The post's comparison is useful: an AI-created visual in a campaign may need a different assessment from a realistic synthetic video of a person making a statement ^[16]. Article 50 does not prohibit AI use, and sets a transparency baseline around content that can mislead people about what is real, who said something, or whether an event occurred ^[17].

Two of the Commission's outputs are optional; the duty underneath them is binding. The Code of Practice for AI-generated content is voluntary, published to help providers and deployers meet the binding Article 50 requirements ^[5][6]. The icon set is free, three icons plus variations, and it sits alongside the required disclosure; on its own it does not make content identifiable in a machine-readable way ^[7][8].

The build work is the origin record. Content usually comes from a mix of internal tools, agencies, freelancers, stock assets and social platforms, which makes it easy to lose track of whether the final asset was generated, edited or materially manipulated by AI ^[12]. The recommended list starts with an inventory of AI content tools used by employees, contractors and agencies, then recording origin and editing history for anything that may be published externally ^[13]. It adds a review step before publication for realistic synthetic media and relevant public-interest copy ^[14], and asks teams to establish who decides whether a disclosure is required ^[15]. Agencies and external creators may need clearer delivery requirements so that clients can make informed publishing decisions ^[19].

Material placed on the market before the rules took effect should be reviewed before December 2026 ^[10], about four months after the obligations landed ^[18]. That review runs on the same origin record, because only assets someone can identify as AI-generated can be re-labelled ^[12].

I would write the mark at generation, in the tool that produced the file, since every later export is a transcode with no reason to know. The dev.to write-up is a practitioner summary rather than the regulation: it notes that the text obligation is subject to exceptions, and names neither the exceptions nor a format for the machine-readable mark ^[20]. The Commission's guidance addresses placement, accessibility and interoperability ^[9].