AI Avatars and the EU AI Act: What Organisations Need to Prepare

Insights / AI policy AI Avatars and the EU AI Act: What Organisations Need to Prepare Article 50 changes more than the disclosure line. It changes how organisations need to manage reusable digital identities. Olena Soinikova · Research & field notes · 30 September 2026 On this page The AI Act regulates the output, not the word “avatar” A compliant platform does not settle the publication decision Consent belongs to another layer of the system The difficult part appears after the first video Governance starts before generation What NGOs should pay particular attention to Article 50 of the EU AI Act has applied since 2 August 2026. For organisations producing AI video, one consequence is immediate: some decisions that used to sit entirely inside a communications workflow now have a compliance dimension. The visible part is disclosure. The more interesting part is what happens behind it.   I work with AI content production across NGOs, EU-funded projects and SMEs through Soinikova Digital, with a team of four AI content creators. The production logic has changed quite quickly. An avatar created for one video rarely has to remain one video. The same digital presenter can return in another language, appear in social media edits and later be considered for training or an interactive service.   At that point, the organisation is managing a reusable digital identity.   This is where Article 50 becomes more useful as an operational signal than as another regulation to summarise. The AI Act regulates the output, not the word “avatar” There is no separate legal category for an “AI avatar” in the AI Act. For organisations publishing avatar video, the relevant analysis can instead lead to the definition of a deepfake in Article 3(60). The definition covers AI-generated or manipulated image, audio or video that resembles an existing person, object, place, entity or event and can falsely appear authentic or truthful. That produces quite different situations inside the same production software.   A clearly synthetic presenter may sit differently from a digital version of a real director delivering a script that the director never recorded. Voice cloning matters because audio is covered by the definition. AI dubbing can create a more difficult case when the original speaker appears to deliver words or speak a language that was generated later.   I find this distinction more useful than asking whether AI was involved. In most contemporary production workflows, that question tells very little.   What matters is what was generated, who or what it resembles and how the audience is likely to understand what it sees. A compliant platform does not settle the publication decision Article 50 divides responsibility between different actors. Under Article 50(2), providers of generative AI systems have obligations concerning machine-readable marking and the detectability of synthetic outputs. Article 50(4) creates a different obligation for deployers where image, audio or video constitutes a deepfake: the artificial generation or manipulation has to be disclosed. Article 50(5) sets requirements around how and when that information reaches people. For a communications team, the practical distinction is straightforward. A technical marker embedded by a platform and a disclosure understood by the audience perform different jobs. This becomes less tidy once a real production process is involved. A script may be approved inside an NGO. An external creator generates the video. Somebody else produces a translated version. The communications manager publishes it, and three weeks later a consortium partner cuts 20 seconds from the original and uploads the extract to another social channel. There is no single “AI compliance button” travelling through that chain.   The organisation therefore needs to know where its own responsibility begins and what happens to the relevant disclosure when content moves. Consent belongs to another layer of the system A real person’s avatar introduces a second issue that is easy to merge with transparency.   An employee, trainer or external expert may have agreed to the creation of an avatar. That permission can matter for GDPR, contractual arrangements and national rules concerning image and voice rights. It does not remove the Article 50 analysis.   Consent is absent from the Article 3(60) definition of a deepfake. Article 50(4) does not contain a general exemption for an avatar created with the represented person’s agreement.   Operationally, I separate these questions. One concerns the organisation’s permission to create and use a digital representation. The other concerns what an audience needs to know when it encounters the resulting material.   The separation becomes important when the first production brief changes.   A person may approve an English training video. Six months later, the organisation wants a Greek version, a fundraising edit and several short videos for social media. Technically, producing them may take very little effort. The original permission may tell the organisation considerably less about whether those uses were actually authorised.   This is one reason AI avatars should stop being treated as ordinary media files once they become reusable. The difficult part appears after the first video Consider a project that records an expert speaking in English.   Later, AI is used to create a Danish version. The translated voice is synthetic and the lip movements are adjusted to make the delivery appear natural. The Danish video is then shortened for social media and distributed through project partners.   Several decisions now sit inside what initially looked like localisation.   The organisation needs a basis for using the person’s likeness and voice in the translated version. It also needs to examine the final output: if the expert appears to have personally recorded the Danish statement, the deepfake definition becomes relevant.   Distribution creates another problem. A disclosure placed next to the original video on a project website may disappear when a partner downloads the file. A social platform can separate the video from its original description. An edit can remove the opening seconds where information was presented.   The published version is therefore what matters operationally, rather than the master file sitting in a