Does this cut need an Article 50 label?

Walk the cut through this tree. If a generated or manipulated image, audio, or video resembles a real person, place, or event and would look authentic to the person who hits it, Article 50 wants a visible or audible label at first exposure. If it is standard assistive editing, evidently artistic in a way that still discloses existence, or generated and published before 2 August 2026, the tree stops. This is a yes/no walk, not an essay on who is the provider.

The Commission's FAQ on Article 50 (last updated 24 July 2026) is the source for the deepfake definition, the first-exposure rule, the artistic carve-out, the assistive-editing exemption, and the non-retroactivity line. Bird & Bird's 2026 insight on the final guidelines (adopted 20 July 2026) adds the advertising and film examples the FAQ compresses. Stop when a node says stop. Do not collect extra badges to feel safe.

Start with the authenticity test

Article 3(60) of the AI Act defines a deepfake as AI-generated or manipulated image, audio, or video that resembles existing persons, objects, places, entities, or events and would falsely appear to a person to be authentic or truthful. The FAQ turns that into three cumulative criteria. All three have to be true.

1. Resemblance. There is a high level of similarity between the content and the simulated subject. A loose sketch of a street is not this node. A photoreal talking head is.

2. Existing. The simulated person, object, place, entity, or event exists, can plausibly exist, or could have plausibly existed. Bird & Bird records that the Commission did not narrow "existing" to people who already have a passport. A photoreal portrait of an invented person can still sit inside the definition. Digital replicas, realistic AI-generated avatars, and a person's image, voice, behaviour, or performance are in the "persons" bucket.

3. False appearance of authenticity. The content could deceive or mislead a person about whether it is authentic or truthful. Intent to deceive is not required. Photorealism makes this more likely; it is not decisive.

The FAQ lets you weigh the level of resemblance, the substantive message, the intended and foreseeable deployment context, and the intended and reasonably foreseeable audience. If that audience does not expect the content to be authentic in that context, the clip may fall outside the definition even though it is synthetic. The Commission's own film example: AI-generated background scenes, special effects, and technical pre- and post-processing as part of standard movie production are not likely to make the work falsely appear authentic. Fully generated actors, digital replicas of real or deceased performers, de-aging, and simulated performances are likely to do so.

Run the test as a sentence at export:

Would a person in the actual foreseeable audience take this picture, this voice, or this event as a real recording?

• Yes → keep walking.
• No, because it is obviously fantastical, animated, or a graphic nobody would file as documentation → stop. Article 50(4) is not asking for a deepfake label on a unicorn. Other rules may still apply.
• Unsure, and the audience includes children or other vulnerable groups → the guidelines use a lower threshold. Keep walking.

A real product against a generated backdrop is the advertising example Bird & Bird flags. If the advertisement does not mislead about the product's actual representation, characteristics, or use, the Commission treats that background as acceptable. If the generate makes the product look different from, or better than, reality, it is a deepfake. The product shot is the tell, not the sky.

This tree does not decide whether you are a provider. If you prompted a hosted model and you are about to publish, you are the person holding the cut. Provider or deployer is the sorting post. Leave it closed while you walk these nodes.

First exposure, not a buried toggle

If the authenticity test fired, the next node is placement.

The FAQ: deployers must disclose deepfake content to a natural person upon first exposure at the latest. The disclosure must be clear and distinguishable, understandable and perceivable by natural persons — visible or audible labels — without any specific technical tools or dedicated extra actions. You cannot simply rely on the machine-readable marking the provider embedded under Article 50(2).

Walk it:

Can a person perceive that this was generated or manipulated at the moment they first hit it, without expanding a caption, opening a menu, or running a detector?

• Yes, because a line or an official EU icon sits on the picture for long enough to read, or an audible line sits on the first seconds of the soundtrack → this node passes. You still need the wording to be true. The disclosure template is the copy. This tree only asks whether the person can perceive it.
• No, because the only mark is a Studio checkbox that appears after "show more," a hashtag, a C2PA manifest, or a two-frame flash → fail. Move the disclosure to first exposure and re-run.
• The clip is audio-only → audible disclaimer. A visual icon the listener never sees is not first exposure.

Paul, Weiss (4 August 2026) records the guidelines' examples of first interaction or exposure: launching a conversation with a chatbot, or the start of a deepfake video. A single up-front notice may not suffice where people may not see the start — a podcast, a mid-roll clip, a video that autoplays from a timestamp. If the foreseeable watch does not include your open, the line has to live where the watch actually begins.

Platform toggles are a different tree. TikTok's built-in AI label, YouTube's altered-content disclosure, Meta's C2PA/IPTC path and self-disclosure are policy. They can help. They do not replace the perceivability test. If the platform badge is easy to miss on mobile, burn the line into the file anyway.

Article 50 is the name of the legal duty. The AI content label is the platform toggle. Walk this node for the legal duty. Then set the toggle so the platform does not suspend you for a different reason.

Carve-outs that actually stop the tree

Three stops are real. They are narrower than a moodboard.

Assistive editing. The FAQ puts standard assistive editing outside the marking obligation, and the guidelines give examples of what stays "standard" versus what goes beyond. Grammar and spell-checking, format conversion, minor cropping, colour, denoise, captions on a real recording — that cluster is the exemption. A face that was never in the room is not an edit. A voice that impersonates a person who did not speak is not an edit. If you only graded a camera original, stop. If you swapped the face, keep walking.

Artistic, creative, satirical, fictional works. The FAQ: some deepfakes are part of evidently artistic, creative, satirical, fictional, or analogous works or programmes. For those, the transparency obligation is limited to disclosure in an appropriate manner that does not hamper the display or enjoyment of the work. The load-bearing word is "evidently." The artistic carve-out is the boundary post. This tree only needs the stop rule: the disclosure still exists; it may sit in an end card, a credit, or accompanying materials so it does not spoil the work. Deepfake content that is exclusively informative or commercial cannot use the quieter placement. Where content combines informative and creative characteristics, Bird & Bird records that the informative character prevails. A UGC-style ad built to look like a real customer is a poor fit for the carve-out. A short film that is obviously a short film is the intended shape.

Generated and published before 2 August 2026. The FAQ (24 July 2026): content generated prior to 2 August 2026 does not need to be labelled retroactively. The Commission encourages voluntary labelling of older published work. Paul, Weiss (4 August 2026) distinguishes generation from, in the case of Article 50(4), publication. The conservative producer rule: if you generated and published before 2 August, stop. If you generated before and you are publishing after, keep walking and label. Public-interest text is cleaner — publication date is the cut. Do not treat a July render sitting on a drive as already "published."

A fourth stop people invent is "I used AI only for the script." A script is not a deepfake. The picture and the voice are. If the face on screen is you and the words were drafted by a model, this tree does not fire on the script. If the face is generated, the tree already fired at node one.

Public-interest text is a parallel tree, not this one. The FAQ's three criteria for that duty: published, informative to the public, on matters of public interest (politics, public health, consumer safety, the environment, financial or scientific developments, and the rest of that list). Human review or editorial control — substance, not spell-checking — removes the text label. Do not drag a video cut through the text exemption. Do not drag a news brief through the deepfake nodes.

What this tree is not deciding

It is not deciding who built the generator. It is not deciding whether TikTok's Creativity Program will pay the post. It is not deciding whether YouTube's inauthentic-content policy will like the channel. It is not handing you paste-ready copy. Those pages exist. This page is the walk from "I have a file" to "does Article 50 want a perceivable mark on this file."

Work the nodes in order. Do not skip to the artistic carve-out because the brand is "creative." Do not skip to assistive editing because a model sat somewhere in the timeline. Do not skip to the platform toggle and call the legal tree done.

If the authenticity test fired, first exposure is missing, and no carve-out stopped you, the answer is yes: this cut needs an Article 50 label. Put it on the picture or on the first seconds of audio. If you generate in Versely's AI video generator, add the overlay before export rather than hoping an upload form will carry it. A request like add the text "This video is AI-generated" as a small overlay in the lower-left for the full clip is an edit step, not a legal conclusion. If you need an account to run that step, sign up.

If every node said stop, do not paste a label on work the Act does not cover just to feel safe. False labels train people to ignore real ones. The tree is a filter. Use both exits.

FAQ

Does a generated background behind a real product need a label?

Not automatically. Bird & Bird's reading of the final guidelines (20 July 2026) treats a real product against an AI-generated background as acceptable where the advertisement does not mislead about the product's actual representation, characteristics, or use. If the generate makes the product look different from, or better than, reality, treat it as a deepfake and keep walking. The product is the test, not the wallpaper.

If I ticked YouTube's altered-content box, did I finish the tree?

No. That box is YouTube policy. Article 50 wants a disclosure a person can perceive without a dedicated extra tap. A note that lives in an expanded description is a weak first-exposure mark on a phone. Burn the line into the file or speak it on the open, then keep the Studio checkbox so you are not fighting two regimes at once.

What if I am unsure at the last node?

Label. The cost of an extra perceivable line on a photoreal clip is a line. The cost of an unlabelled realistic deepfake is a failed duty. If the only reason you are unsure is "maybe the artistic carve-out applies," read the carve-out post before you quiet the placement. Uncertainty about authenticity is not a carve-out. It is a reason to use the louder line.