Labelling AI content: the EU AI Act and platform rules

Short answer: Article 50 of the EU AI Act sets two main duties: providers of generative AI systems must mark outputs in a machine-readable way, and deployers who publish deepfakes must disclose that the content is artificial. The article has applied since 2 August 2026, with a transitional period until 2 December 2026 for the marking duty for systems placed on the market before that. Platforms (YouTube, Meta, TikTok) have their own label rules too. This article is a summary for creators and business owners; it is not legal advice.

What does Article 50 of the EU AI Act require?

Article 50 regulates transparency obligations. In outline:

• 50(1): Providers of systems that interact directly with people must ensure people know they are dealing with an AI system, unless that is obvious from the context.
• 50(2): Providers of generative AI systems that produce audio, image, video or text must mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated. The Commission expects solutions to be effective, interoperable, robust and reliable as far as technically feasible.
• 50(3): Those using emotion recognition or biometric categorisation systems must inform the people exposed to them.
• 50(4): Deployers that use systems to generate or manipulate deepfakes and publish the content must disclose that it is artificial. A similar disclosure is needed for AI-generated text published to inform the public; there is an exception where there is human review and editorial responsibility.
• 50(5): The information must be given clearly and distinguishably at the latest at the first interaction or exposure.

The obligation is lighter for evidently artistic, creative, satirical, fictional or analogous works: disclosing the existence of the content in an appropriate manner that does not hamper the display or enjoyment of the work is enough. The Commission's FAQ page also mentions specific regimes for such works.

Who has to do what?

Party	Article	Duty
Provider of a generative AI system	50(2)	Mark output in a machine-readable way, make it detectable
Deployer publishing a deepfake	50(4)	Disclose that the content is artificial
Those publishing AI text to inform the public	50(4)	Disclose it is artificial (if no editorial review)
Provider of an interactive system such as a chatbot	50(1)	Inform that people are dealing with an AI

A person who uses a tool and shares on their own account is mostly the addressee of the disclosure duty (as a deployer) rather than the marking duty (as a provider). This distinction needs legal assessment in a real case.

Timelines: when does it apply?

The Commission's FAQ pages are clear: Article 50 applies from 2 August 2026; this date holds for systems under 50(2) and 50(4). For systems placed on the market before 2 August 2026, there is a transitional period until 2 December 2026. This was introduced by legislative changes on AI. Also, outputs made available before 2 August 2026 do not need to be marked or labelled retroactively.

The Commission published a voluntary code of practice to help comply with these obligations (June 2026). Section 1 covers marking and detection for providers, and Section 2 covers labelling of deepfakes and published text for deployers. Signing is voluntary; those who do not sign must show compliance by other means. The Commission page also says it gives guidance on label design and placement and offers an optional EU icon.

What do platforms require?

The law is one thing, platform rules another. The following is summarized from the platforms' own official pages and may change over time:

Platform	What is expected?	How does the label look?
YouTube	Realistic content that is AI-generated or meaningfully altered must be disclosed; for example, making a real person appear to say something they did not, altering footage of a real event or generating a realistic scene that did not occur. Minor aesthetic edits such as beauty filters, colour correction and captions, and non-realistic content do not need disclosure.	The "AI use" setting in YouTube Studio during upload; a label in the player or in the expanded description
Meta	Meta applies an "AI info" label when it detects industry-standard AI indicators or when people disclose at upload that content is AI-generated	An "AI info" label; a more prominent label for high-risk content
TikTok	The help page explains which AI content creators should label and how content carrying C2PA content credentials is labelled automatically	See the current version of the page for details

Practical upshot: read the current help page of the platform you will post on and use its disclosure setting; many platforms apply consequences such as warnings or removal for breaking the rule, so find the details on the relevant page.

What are C2PA and content credentials?

C2PA is a coalition developing open technical standards for content provenance and authenticity. Content Credentials is a system built on that standard: a mark appears on the content, and clicking it shows how the content was made and its editing history. The page acknowledges that bad actors will keep trying to label synthetic content as authentic; the goal is to give good-faith creators a way to show their content is authentic.

Producers use other marks too. For example, Google states that videos made with Veo are marked with SynthID, its technology for watermarking and detecting AI-generated content. As general background, file metadata can be lost through re-saving and screenshots, so marks are a signal, not conclusive proof. You should also add a visible label or disclosure text to say the content is artificial.

A practical checklist for creators and businesses

1. Decide the type of content: is it realistic content that resembles a real person, event or place, or clearly imaginary?
2. If realistic, disclose: a visible label ("Generated with AI") or the platform's disclosure setting.
3. Check platform rules: read the current pages of YouTube, Meta and TikTok.
4. Do not strip tool marks: do not try to remove content credentials and watermarks.
5. Keep records: save what you made, with which tool and when.
6. Get people's consent: permission is essential for real people's likenesses. See our article What is a deepfake?.
7. For text: if you publish AI text informing the public, consider taking on editorial review and responsibility.

Example disclosure: "This image was generated with AI; it does not show a real event."

Example disclosure: "The audio and visuals in this video are an AI-generated recreation."

Example scenarios

The following is a general reading; consult an expert for a definite assessment.

• A clearly imaginary fantasy illustration: If it does not resemble a real person, place or event, it may fall outside the deepfake definition. YouTube's page also says non-realistic content does not need disclosure. Still, check the tool's terms and the platform rules.
• A realistic video of a person saying something they never said: This can fall under the deepfake scope in the law and is a type that needs disclosure in YouTube's examples. A clear label and the person's consent are essential.
• Changing the background in a product photo: YouTube's "minor aesthetic edit" examples list colour and lighting adjustments; replacing the whole background may not count as that. Not misrepresenting the product in a commercial context is a separate matter; if in doubt, label it.
• A blog post written with AI: Article 50(4) asks for disclosure for AI text published to inform the public; there is an exception where there is human review and editorial responsibility. Genuinely review the text before publishing and take responsibility.

Common mistakes

• Saying "labelling is only the provider's job": A deployer publishing a deepfake may have a disclosure duty too.
• Hiding the label or writing it tiny: The information must be clear and distinguishable.
• Saying content credentials are enough: Metadata can be lost; add a visible disclosure too.
• Assuming platform rules do not change: Check the pages regularly.
• Making content of real people without consent: A label does not make up for missing consent.

Summary

• 50(2): providers mark outputs in a machine-readable way; 50(4): those publishing deepfakes disclose.
• Article 50 has applied since 2 August 2026; systems already on the market have until 2 December 2026 for marking.
• The Commission's voluntary code of practice has provider and deployer sections.
• YouTube, Meta and TikTok have their own label rules; read their current pages.
• Marks such as C2PA and SynthID help; add a visible disclosure too.