EU AI Act Article 50: what L&D and media teams have to do about AI-translated video

Who has to mark, who has to disclose, and by when. In plain language, without pretending the answer is always simple.

edon.ai · 26 August 2026 · 9 min read

What Article 50 actually covers

Article 50 of the EU AI Act is the transparency article. It is not about whether you may use AI, and it is not a quality standard. It is about people knowing when what they are watching or listening to was generated by a machine.

It matters for translated video because a soundtrack generated in a synthetic voice is generated audio, and the rules attach to it whether it was made for a marketing campaign, a published show or a mandatory safety module.

Provider duties and deployer duties are different things

This is the distinction that decides what lands on your desk, and it is the one most internal briefings get wrong.

Providers build the generative system and put it on the market. Their duty runs to the file: outputs must be marked in a machine-readable format and detectable as artificially generated. There is a second half people routinely miss, which is that the provider must also make a means of detection available to the people who might encounter the content.

Deployers use the system and put the result in front of an audience. That is you. Your duty runs to the person watching: telling them that what they are seeing has been artificially generated or manipulated, clearly and distinguishably, at the latest at the point they first encounter it.

The consequence is the single most expensive misunderstanding in this area, so it is worth stating flatly: your vendor’s machine-readable marking does not discharge your disclosure duty. A watermark inside a file is not clear and distinguishable to a human being watching a video. If a vendor tells you their marking means you have nothing to do, they are describing their obligation and leaving you holding yours.

The dates, and the transitional period that is not for you

The transparency obligations applied from 2 August 2026. They are in force now, not on the horizon, so if this has been sitting on a roadmap as next year’s project, it has already started.

There is one transitional period and it is narrower than most summaries suggest. Providers whose generative systems were already on the market before 2 August 2026 have until 2 December 2026 to comply with the marking obligation. That is the whole of it: providers only, the marking obligation only, and only for systems that already existed. A system launched after August 2026 gets no grace period, and the deployer disclosure duty had no transition at all.

So if you are a deployer waiting for December, you are waiting for something that was never about you.

When a translated video counts as a deepfake

The disclosure duty attaches to a deepfake, which the Act defines as generated or manipulated image, audio or video content that resembles existing persons, objects, places or events and would falsely appear authentic. In practice the Commission reads that as four things that all have to be true: there is an appreciable resemblance, to something or someone that actually exists, and the result would falsely appear genuine to a viewer.

Two points make this concrete for anyone localising video.

A real person’s voice counts. “Persons” expressly covers personal characteristics including image, voice, behaviour and performance. Synthesising realistic speech in a specific person’s voice is called out as a change that requires marking, and it is excluded from the light-touch exemptions for assistive editing. If your safety officer appears to speak Polish, you are in the territory the rule was written for.

Translated text is treated differently. AI-generated translation of text sits on the exempt list. So translated subtitles or captions on their own are a much lighter proposition than a synthesised voice, and a clearly generic narrator that does not resemble any identifiable person is weaker ground for a deepfake finding than a cloned presenter.

That line, between translating the words on screen and generating a real person’s voice, is the one worth understanding before you choose an approach. It is not a loophole, but it is a genuine difference in what you take on.

Internal training video is in scope too

Here is the part that surprises people, and it is worth reading twice: there is no publication threshold. The duty turns on whether an AI system was used professionally to generate a deepfake. It does not turn on whether the video went out to the public.

Internal corporate training video is in scope. A module that never leaves your learning platform, watched only by your own employees, is covered on the same terms as something you put on a public channel. Any internal briefing that has quietly assumed “this is internal, so it does not apply” needs correcting.

A second trap: the human editorial review exemption, which people reach for a lot, belongs to the AI-generated text limb of Article 50. It does not rescue a deepfake. Having a person check the translation is excellent practice and it is not a disclosure exemption.

What a disclosure looks like in practice

Disclosure does not have to be heavy. It has to be visible to a human, at the point they meet the content.

• Internal training. A line on the module’s landing page in the learning platform, plus a card at the start of the video. Workers should not have to go looking for it.
• A podcast. A sentence in the episode notes and a spoken line in the intro of the translated feed.
• Published media. Your own note in the description, in addition to whatever your platform’s labelling does. A platform label is the platform meeting its obligations, not you meeting yours.

One practical refinement worth building in now: where viewers may not see the beginning, a label only at the start can be insufficient, and the guidance points toward disclosure that is repeated or persistent. If your content is chaptered, deep-linked from a document or resumed mid-way, plan for a marker that survives someone joining late.

Write the wording once, agree it with whoever signs things off, and put it in the template. The teams that struggle are the ones deciding this per video.

What your tools can do, and what stays with you

What a localisation tool can genuinely help with is the record: what was produced, from what source, in which languages, and when. That is the material you will want if anyone asks you to account for a published version a year later.

What stays with you is everything facing your audience: the decision to publish, the disclosure they actually see, and the judgment about whether a given video needs one. No vendor can take those on, and you should be wary of one that says it can.

One thing worth putting in your contracts, especially if agencies or distribution partners sit between you and the viewer: say explicitly who is responsible for the label reaching the audience. In a production chain that runs through several hands, this is exactly the duty that falls down the gap between them.

A checklist you can act on this week

1. List everything you publish or circulate that contains AI-generated or AI-translated audio and video, including internal-only material. Most teams find more than they expected, because marketing usually got there first.
2. For each item, ask whether a real, identifiable person appears to be saying something they did not say. That answers the disclosure question faster than re-reading the article.
3. Agree one standard disclosure wording with whoever owns risk, and put it in the publishing template rather than deciding it each time.
4. Ask your vendors in writing what they mark, what detection they make available, and what records they keep. Keep the answer.
5. Check where responsibility for the visible label sits in your contracts with agencies and distributors.
6. Get speaker permissions in order before the next project rather than after it ships.