Worked example: the campaign video with a synthetic presenter

A marketing team produces a two-minute explainer for a new savings product. The presenter is generated: a synthetic face and a synthetic voice, scripted by the team, rendered by a commercial tool. The tool's vendor confirms in writing that all outputs carry embedded machine-readable provenance marking. The video is scheduled to run on the company website, on social channels, and as a pre-roll advertisement. Marketing asks whether anything else is needed.

Sort the duties before answering

Is this a deepfake for Article 50(4) purposes? It is generated video depicting a person. The team should treat it as in scope and take advice rather than reason its way out, because the consequence of being wrong is a disclosure failure on a published asset.

Does the vendor's marking settle it? No, and this is the question the whole example exists to answer. The vendor is discharging the Article 50(2) provider duty: output marked in a machine-readable format, detectable as artificially generated. That duty is real and it is theirs. The Article 50(4) deployer duty is ours, and the Commission's Article 50 guidelines require a visible or audible label understandable without a detection tool. A viewer cannot perceive embedded provenance metadata, and could not check it if they tried. Nothing has been disclosed to them.

Does the artistic exception apply? No. The limited exception covers content that is evidently artistic, creative, satirical or fictional. A product explainer on a bank's website is none of those. Its whole purpose is to be taken as a straightforward communication from the company.

What the visible disclosure has to look like

Article 50(5) requires the information at the latest at the time of the first exposure, clearly and distinguishably. That has three practical consequences.

At the start, not the end. A caption in the final frame is not first exposure. Most viewers never reach it.

On the asset, not around it. Social platforms strip descriptions, truncate captions and autoplay without sound. A disclosure that lives in the post text is gone the moment the video is reshared. Burn it into the video, and use the platform's own synthetic media flag as well where one exists.

Legible to a viewer, not to a lawyer. "This video features an AI-generated presenter" does the work. A reference to Article 50 does not.

Pre-roll advertising raises the same points more sharply, because the viewer did not choose to watch and may see only the first seconds. Those first seconds must carry it.

What else the team should write down

The script itself is worth a moment. If generated text is published to inform the public on matters of public interest, Article 50(4) requires disclosure of its artificial origin unless it has undergone human review or editorial control with a named person holding editorial responsibility. Here the script was written by the team, so the question does not arise, but the team should record that, because in six months nobody will remember which parts were drafted by a tool.

The record for this asset should name the generation tool and version, what was generated, the vendor's statement about machine-readable marking, the visible label applied and where it appears, the publication channels, and the date. That record is what answers a question from a regulator, a platform, or a customer, and it costs a few minutes at the time and an investigation later.

If the same video goes to China

The Measures for Labelling AI-Generated Content and GB 45438-2025 have been in force since 1 September 2025 and require both an explicit visible label and an implicit machine-readable label in metadata carrying the provider's name and a content reference number. The visible label the team has added for Article 50(4) is a good start and is not automatically sufficient, and the metadata requirement is more specific than the EU's. Check the vendor's marking against what that standard requires rather than assuming it transfers.

The sentence to remember

The provider marks for machines. The deployer labels for people. One does not do the other's work, and the deployer's duty is the one that fails quietly, because nobody complains about a label they never saw.

As of 24 September 2026.