Article 50 disclosure template

Paste these labels. Article 50 of the EU AI Act has applied since 2 August 2026; this page is the copy, the three official EU icons, and where the label sits on TikTok, YouTube, and Meta. It is not a substitute for the law, and it is not legal advice.

A reviewer checking an on-screen AI disclosure against an Article 50 labelling checklist

The Commission's FAQ on Article 50 transparency obligations and the EU icons for labelling AI-generated content are the primary sources. Guidelines were adopted on 20 July 2026. Other Versely posts walk the law: what Article 50 means for sponsored content, who labels, provider or deployer, and the artistic carve-out. This page is the text you can drop on the asset.

Copy-paste labels

Use the shortest line that is true. Put it where a person can perceive it at first exposure — on the file, not only in metadata. The synthetic media disclosure is the legal duty. The AI content label is the platform toggle. They stack. Neither is optional because the other exists.

Video — fully generated

This video is AI-generated.
This video was generated with AI.
Synthetic media — this clip was generated with AI.

Video — partially modified

This video was modified with AI.
This video contains AI-modified footage.
Synthetic media — parts of this video were generated with AI.

Video — voice or soundtrack only

Synthetic media — voices generated with AI.
The voice in this video is AI-generated.
Soundtrack generated with AI.

Image — fully generated

This image is AI-generated.
This image was generated with AI.
Synthetic media — this image was generated with AI.

Image — partially modified

This image was modified with AI.
This photograph was edited with AI.
Synthetic media — this image was altered with AI.

Audio — fully generated

This audio is AI-generated.
This track was generated with AI.
Synthetic media — this audio was generated with AI.

Audio — voice or performance

Synthetic media — voices generated with AI.
This voice is AI-generated.
The speech in this file is AI-generated.

Text — public-interest material, no human review

This text was generated with AI.
This article was generated with AI and has not been reviewed by an editor.
AI-generated text — not reviewed.

Text — you did review it (say so; the label may then be unnecessary under Article 50)

Drafted with AI; reviewed and approved by [name / desk].
AI-assisted draft; a human editor checked the facts and takes responsibility for publication.

Spell-checking is not review. The Commission's FAQ is explicit: human review means a person with relevant knowledge examined the substance; editorial control means someone with authority to approve, alter, or reject the content on substantive grounds, including fact-checking. Grammar passes do not count. If you only ran a linter on the sentences, you still need the generated-text label for public-interest material.

Keep a house style. One channel, one wording, every time. "This video is AI-generated" in the same corner, same size, same duration, is easier for a viewer to learn than a witty new line on every post. Witty lines fail first-exposure tests. The duty is that a natural person can perceive the disclosure without a special tool and without a dedicated extra tap.

Burn the line into the file when the destination re-encodes. A caption that lives only in an upload form dies when the clip is downloaded and re-uploaded. In Versely, a request like add the text "This video is AI-generated" as a small overlay in the lower-left corner for the full clip runs the text overlay step. Iterate on a preview; pay credits on the export. The overlay travels with the MP4.

Do not paste a label on work the Act does not cover just to feel safe, and do not skip it on work that looks real. Standard assistive editing — colour, crop, denoise, captions on a real recording — is exempt. A photoreal talking head that was never in the room is not. If the clip would falsely appear authentic, use a fully-generated or partially-modified line from the lists above. If you are unsure, use the louder line.

The three EU icons

The Commission published three icons for deployers. Using them is optional. Labelling is not. The icons do not, by themselves, establish legal compliance. Download the files from the Commission; do not redraw them.

The pack is on EU icons for labelling AI-generated content (page last updated 10 August 2026). Official downloads:

Four visual variations exist for each icon: black, white, black at 50 percent transparency, white at 50 percent transparency. The Commission reports that user-testing improved when the basic icon was accompanied by a text label.

Basic icon. Use when AI was involved in creating a deepfake (image, audio, video) or published text, or when you pair the mark with a custom text label or an interactive second layer. Commission example: a deepfake video with the text "voices generated with" followed by the basic icon. That is the same job as the paste line "Synthetic media — voices generated with AI."

Fully AI-Generated. Use when the entire deepfake or the public-interest text is fully generated by AI, with no human-created elements and no human editorial control beyond prompting. Commission examples: fully AI-generated deepfake videos of politicians or fictional events; fully AI-composed music or art (those last two may also sit inside the artistic carve-out); AI-generated news summaries.

Partially AI-Modified. Use when pre-existing, human-made content was partially modified with AI and the result is a deepfake or public-interest text. Commission examples: a face in an authentic photograph swapped for a politician's face; authentic photographs of an empty apartment furnished with AI.

The basic icon is the one designed to carry a second, interactive layer — further information that a viewer can open. If you implement that layer, the icon should show that more information exists, and the second layer must be navigable with assistive technologies. User-testing favoured "basic icon + plain words" over the icon alone. Pair the mark with one of the paste lines. Do not invent a fourth icon. Do not ship a custom SVG and call it the EU mark.

Placement rules from the Commission page, summarised, not complete:

Clearly perceivable and distinguishable at the latest at first exposure.
Placed where no intervening overlay covers it.
Directly embedded in the deepfake or published text (except for creative works), unless an equivalent user-interface overlay exists. The icon must remain visible when the content is reshared or downloaded.
Large enough to see. Accompanying text in plain language; the only abbreviation the Commission flags as acceptable is "AI."
If the disclosure is timed, leave it up long enough for someone with slower processing to read it.
Alt text or ARIA labels so assistive technology can announce that the content is AI-generated or manipulated.

Signatories of the Code of Practice on marking and labelling of AI-generated content must follow the Code's placement specifications. Everyone else may still use the icons freely; the Commission does not require attribution. Using the icons is not a claim that you signed the Code.

Machine-readable marking is a different artefact. Providers of generative systems must mark synthetic audio, image, video, and text so it is detectable. That is content credentials / C2PA, watermarks, or another embedded signal — aimed at a detector, not a scrolling viewer. A green provenance tick is not your on-screen label. Social platforms re-encode on upload; a manifest often dies there. Burn the human-readable line anyway.

Where to put the label

Article 50 asks for a visible or audible disclosure at first exposure. Platform toggles are extra, and they differ. The label is not a ranking penalty on TikTok, YouTube, or Meta; unoriginal, templated content is. Disclose, then worry about whether the clip is actually yours.

TikTok. Use the built-in AI label. TikTok's Creativity Program pays labelled AI content at the same rate; undisclosed AI can mean removal or program suspension. Self-apply the label rather than waiting for C2PA detection to do it for you. Detection can still fire if you skip the toggle, which means you lose the wording. Put one of the video paste lines in the first lines of the caption as well, and burn the same line into the file so a download still carries it. TikTok has said adding the label does not affect distribution. The Act does not care about distribution; it cares that a person can see or hear the disclosure when they first hit the clip.

YouTube. Use Studio's altered-content disclosure for realistic synthetic media of real people, or a realistic scene that did not happen. The platform label typically appears in the expanded description. That is easy to miss on mobile, so do not treat the checkbox as the only first-exposure mark. Burn "This video is AI-generated" (or the modified variant) on the picture, or speak "This video is AI-generated" on the first seconds of the soundtrack. YouTube's inauthentic-content policy bans mass-produced, repetitive, templated uploads — not AI as such. AI content stays monetizable when it carries the creator's own vision. The disclosure and the originality test are two different gates.

Meta (Facebook, Instagram, Reels). Labels are required. Meta relies on C2PA/IPTC signals and self-disclosure. Reels bonuses depend on disclosure. Facebook may apply an "AI info" / "Made with AI" label via metadata, classifiers, or a toggle you set. Instagram has also tested a profile-level "AI Creator" statement; that is an account-level badge, not a per-post substitute for Article 50. Self-disclose on the post, burn the paste line into the video or image, and do not assume a classifier will word it correctly. A photoreal Reel of a person who was never filmed is a deepfake problem first and a Reels-bonus problem second.

On every platform, first exposure means the viewer does not have to expand a caption, wait until the last second, or catch a two-frame flash. A lower-third that sits for the duration of a short is the boring version that works. An audible line at the open works for audio-first posts and for viewers who cannot see the overlay. Use both when the clip is photoreal and short.

A clip from Veo 3.1 or any other generator is still your publication when you post it. The model page is not a label. The AI video generator will give you the file; the overlay step puts the words on it. If you need an account to run that loop, sign up and treat the disclosure as part of the edit, not as a checkbox you remember at upload.

Fictional, satirical, or evidently artistic work may use a quieter placement — an end card, a first-line description, a standing channel credit — so the disclosure does not spoil the work. The load-bearing word is "evidently." A UGC-style ad built to look like a real customer is a poor fit. The artistic carve-out is the boundary post. This template still gives you the words; you only change how loud they are.

What this template does not cover

Four areas sit inside Article 50. This page only hands you copy for generated content and for deepfakes / public-interest text.

Direct interaction. Chatbots, agents, and avatars that talk to a person must disclose that the person is talking to AI, unless that is obvious, from the first interaction. That is a product-UI problem, not a YouTube overlay.
AI-generated content (provider marking). The machine-readable mark in the file. Providers of systems already on the market before 2 August 2026 have a grace period to 2 December 2026. Watermark-detection interoperability runs to 2 February 2027. If you only prompt a hosted generator, this duty is usually not yours; you still cannot hide behind a missing watermark.
Emotion recognition and biometrics. Deployers must inform people exposed to those systems. That is a notice at the camera, not a TikTok caption.
Deepfakes and AI text on public-interest matters. This is the deployer label this template is for: visible or audible, at first exposure.

Dates that actually move a calendar:

20 July 2026 — Commission guidelines adopted.
2 August 2026 — Article 50 applies. Deployer labelling of deepfakes has no grace period.
2 December 2026 — provider marking grace for systems that were already on the market.
2 February 2027 — watermark-detection interoperability backstop.

Content generated and published before 2 August 2026 is not retroactive. Content generated before 2 August 2026 but published after that date is covered. If you are sitting on a drive of unreleased Veo takes from July, the publication date is the one that counts. The Commission encourages voluntary labelling of older published work; it does not require you to go back through the archive.

Who pastes the line: the deployer, meaning the party that used the system under their authority and then put the content in front of people. Personal, non-professional use is outside the Act. Regular economic benefit — a creator who is paid, a brand, an agency — is professional use. Employees of a company are not separate deployers; the legal person is. Provider or deployer is the sorting post. If you prompt a hosted model and publish the output, you are the person this template is for.

Fines for relevant breaches can reach €15 million or 3 percent of worldwide turnover, with proportionality for SMEs. For most independent producers the near-term risk is a platform strike or a client refusing the cut. The ceiling is why buyers now ask to see the overlay in the review link.

Do not use this page as a licence to skip advertising disclosure, likeness releases, or platform originality rules. A labelled fake testimonial is still a fake testimonial. A labelled clip that is the same six-second template as a thousand other accounts is still inauthentic content on YouTube. The paste line answers "was this generated or manipulated." It does not answer "is this yours" or "were you paid."

FAQ

Does every AI clip need one of these labels?

No. Article 50(4) targets deepfakes — generated or manipulated image, audio, or video that resembles existing people, objects, places, entities, or events and would falsely appear authentic — plus AI text on public-interest matters that was not substantively reviewed. Standard assistive editing is exempt. Obviously fantastical work may fall outside the deepfake definition. When the clip looks like a real recording of a real person, paste the line.

Is TikTok's or YouTube's AI toggle enough on its own?

No. Platform labels are policy. Article 50 is a visible or audible disclosure a person can perceive without a special tool. A Studio checkbox that only appears after "show more" is a weak first-exposure mark. Burn the paste line into the file, set the toggle, and keep the caption consistent.

What if I made the file before 2 August 2026?

If you also published it before that date, the Act is not retroactive. If you generate it before 2 August 2026 and publish it after, it is covered. New posts of old renders get the same overlay as anything you generate today.