AI Labelling Requirements in 2026: Which Half of Article 50 Is Actually Yours

You read the first three paragraphs of an AI Act explainer, hit the line about model providers having to embed machine-readable marks, and closed the tab. Not my problem. That's an OpenAI problem.

You relaxed at the wrong sentence. Article 50 has four operative paragraphs and two different kinds of duty-holder — and one of them is defined as the person who actually deploys the system and publishes the output. That is you.

Four paragraphs, two duty-holders — the split everyone gets wrong

Here is the structure, with the operative language quoted (the consolidated text is published in the EU's Official Journal; the article text is also mirrored publicly):

• 50(1) — providers. Systems intended to interact directly with people must be built so people know "they are interacting with an AI system" — "unless this is obvious."
• 50(2) — providers. Systems generating synthetic audio, image, video or text must ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated." This is the tool maker's job, not yours.
• 50(4), first limb — deployers. Anyone using an AI system that generates or manipulates image, audio or video content constituting a deep fake must "disclose that the content has been artificially generated or manipulated." This one is yours.
• 50(4), second limb — deployers. Text generated or manipulated and published to inform the public on matters of public interest must be disclosed.
• 50(5) — everyone. Disclosure must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure."

One line to remember the split: the model writes it into the file, you say it to the human. Doing one does not cover the other.

The three exemptions, read precisely

Every summary lists the same three carve-outs, and every summary makes them wider than they are.

• "Assistive editing is exempt." The actual limiter is that the function must not "substantially alter the input data" or its semantics. Colour grading, denoising and cropping usually sit inside that. Generative fill, background replacement and object removal are much harder to argue.
• "Comedy and fiction accounts are exempt." This is the most over-read line in the whole article. Where content is evidently part of an artistic, satirical or fictional work, the obligation is limited — but disclosure is still required, in a manner that does not spoil the display or enjoyment of the work. It is a change of format, not a free pass.
• "Someone reviewed it, so it's exempt." This applies to text only, and requires human editorial review with a person or entity holding editorial responsibility. Reading your own draft once before posting does not create editorial responsibility in the legal sense.

There is a fourth limiter that almost no one mentions: 50(4)'s first limb covers deep fakes, a defined term — not "everything a model touched." Reading it as "all AI content must be labelled" is both wrong and counterproductive, because it makes you numb to the cases that genuinely do trigger it.

We are deliberately not quoting any fine amounts. Penalty mechanics and national enforcement practice were not verified for this piece, and a confident number you can't act on is worse than no number.

On scope — it governs the EU market, and that "only" is wider than it sounds. The obligation attaches to content placed in front of EU users. For anyone running international accounts, the practical problem is that you do not decide who sees your post — the recommendation system does. A Reel or a TikTok gets distributed by an algorithm that has never read your target-market spreadsheet.

Which makes the sane operating rule not "assess each post for EU exposure" but make disclosure the default. One line of text, or one platform toggle, versus an unbounded downside. That trade is not close.

The second rulebook: platform policy is not the law, and vice versa

This is where most creators actually get hurt — not by regulators, but by assuming the two systems are one system. They are not, and they do not even measure the same thing:

• The law asks: was the audience told, clearly and at first exposure?
• The platforms ask: did you use our disclosure control, and does your content fall into a category we restrict?

The consequences differ too. A platform will not fine you; it will quietly stop recommending you. Snapchat, for example, has stated that only videos made by real people are eligible for Spotlight recommendations, while still allowing its own AI editing tools — a distribution rule, not a legal one. We keep a running comparison in AI content disclosure rules by platform, and the Snapchat wording specifically in Snapchat's Spotlight AI policy.

The layer you cannot see — and why deleting the badge does nothing

Article 50(2)'s machine-readable marking is not a visible watermark. It is metadata riding inside the file. China's parallel rulebook is structured identically: an explicit label a user can perceive, plus an implicit label written into the file's metadata carrying the content's attributes, the service provider's name or code, and a content ID.

The practical consequence is the one that catches people out: cropping the corner badge and re-exporting removes the part a human sees and usually leaves the part a machine reads. And in the Chinese framework, maliciously deleting, altering, forging or concealing that label is itself prohibited — as is providing tools or services that help others do it.

If you got here searching for how to strip a label, the honest answer is that the question points the wrong way. What the label actually breaks — and what your options are — is covered in why a Content Credentials label appeared and what you can legitimately do about it. On the separate question of who owns AI output, see is AI-generated video copyrighted.

A five-minute compliance pass you can actually run

• Sort your content by whether a viewer would take it as a record of something real. That question, not "how much AI did I use," is the one both rulebooks are circling.
• Make disclosure a publishing step, not a judgement call. The moment it becomes a per-post decision, it will be skipped on the night you are tired.
• Keep the file intact. Re-encoding to "clean" a file costs you quality and buys you nothing.
• Handle commercial content separately. Paid partnership disclosure is a third, unrelated obligation — see the paid partnership label on every platform.
• Write down which rulebook each of your habits satisfies. Most teams discover they have been covering one and calling it two.

What "clear and distinguishable" probably means in practice

Paragraph 5 is short and does a lot of work: disclosure must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure." Two words in there decide most edge cases.

"Distinguishable" means the disclosure has to be separable from the content — not woven into the creative such that a viewer would read it as part of the fiction. A line of dialogue in your video saying "this is fake!" is content. A caption line or platform label is disclosure.

"At the latest at the time of first exposure" means it cannot live where the viewer arrives second. This is the same structural problem TikTok addressed from a completely different direction in its branded content rules, which require that what you are promoting be "sufficiently clear, without requiring viewers to access your profile page or any links." Two different rulebooks, written by different people, arriving at the same conclusion: obligations that live in your bio do not travel with your content.

The practical translation for short video is unglamorous: put it where the video is, not where your profile is. A first-frame or on-screen line, or the platform's own label, satisfies both readings. A note in a link-in-bio satisfies neither.

There is one more asymmetry worth internalising. Under-disclosing is a compliance risk with an unbounded ceiling. Over-disclosing costs you a line of text, and — as far as any platform has published — nothing else. When you can't decide, the expected values are not symmetrical, and the cheap option is also the safe one.

FAQ

Do I need to label an AI-written caption on a normal lifestyle post?

Under Article 50(4)'s text limb, the trigger is text published to inform the public on matters of public interest. A caption on a lifestyle post is usually outside that. But platform policy is a separate question, and commercial or health-adjacent claims can pull you back in through other rules entirely.

Does an AI voiceover count as a deep fake?

The distinguishing question is whether it imitates a specific real person. A generic synthetic voice reading your own script is a different risk profile from cloning an identifiable individual's voice — the latter adds personality and voice rights on top of any disclosure duty.

If I flip the platform's "AI-generated" toggle, am I done?

You have satisfied the platform. Whether you have satisfied the law depends on whether that disclosure is clear, distinguishable, and present at first exposure. In most cases the toggle gets you there — but treat it as the necessary step, not the complete one.