What EU AI Act Article 50 Requires When You Post AI Content

What is EU AI Act Article 50?

Transparency is the whole job of EU AI Act Article 50, the article of Regulation (EU) 2024/1689 that tells providers of AI systems to mark what those systems generate and tells deployers to disclose deepfakes and certain AI-written text. It sits in Chapter IV, "Transparency Obligations for Providers and Deployers of Certain AI Systems", and has seven paragraphs. Two of them decide almost everything for a social media team: Article 50(2) and Article 50(4).

Article 50 is not a high-risk rule. Chapter III covers a defined subset of AI uses; Article 50 covers any AI system used in one of the four situations it names. A brand with no high-risk AI still lands inside it.

Is Article 50 in force yet?

Yes. Article 113 sets the general application date as 2 August 2026, and Chapter IV is not carved out of it by any of the exceptions in Article 113 points (a) to (d). Article 50 binds today.

Most explainers get this wrong: they were written while the date was still ahead and still say the obligations "will apply".

Date	What starts to apply	Basis
1 August 2024	Entry into force	Article 113
2 February 2025	Chapters I and II	Article 113(a)
2 August 2025	Chapter III Section 4, Chapters V, VII and XII, and Article 78	Article 113(b)
2 August 2026	General application date, which brings in Chapter IV and Article 50	Article 113
2 August 2027	Article 6(1) and its corresponding obligations	Article 113(c), as amended

Are you a provider or a deployer under Article 50?

You are a deployer. Article 3(4) defines a deployer as "a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity". A provider, under Article 3(3), develops an AI system or has one developed and places it on the market under its own name or trademark.

A social media team runs someone else's model to make a caption or an image and publishes the result. That is deployment; the company that built the model is the provider. Conflating the two is the most common error in Article 50 coverage, and the two paragraphs demand different things.

	Article 50(2)	Article 50(4)
Binds	Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text	Deployers of AI systems that generate or manipulate content
Covers	Every synthetic output the system produces	Deepfakes in image, audio or video, plus text published to inform the public on matters of public interest
Requires	Marking outputs "in a machine-readable format and detectable as artificially generated or manipulated"	Disclosing "that the content has been artificially generated or manipulated"
Audience for the signal	Machines and detection tools	People looking at the post
In your org chart	Your AI vendor	You
Carve-outs	Assistive function for standard editing, no substantial alteration of input data or its semantics, law enforcement	Law enforcement, reduced duty for evidently artistic or satirical work, human review plus editorial responsibility for the text limb only

What does Article 50(2) require of providers?

Article 50(2) requires that outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated", and that providers "ensure their technical solutions are effective, interoperable, robust and reliable as far as this is technically feasible", judged against the cost of implementation and "the generally acknowledged state of the art".

Recital 133 lists the techniques the legislator had in mind: "watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques". None of that is a visible label. It is provenance data a detection tool reads. Your caption does not satisfy Article 50(2), and your vendor's watermark does not satisfy Article 50(4). Separate signals, separate readers.

What does Article 50(4) require deployers to disclose?

Article 50(4) covers two distinct cases, and they have different shapes.

The first is deepfakes. Deployers of an AI system "that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated". Article 3(60) defines a deep fake as "AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful". A deepfake under the AI Act is visual or audio only; text is not in the definition.

Where the content forms part of an "evidently artistic, creative, satirical, fictional or analogous work or programme", the duty is reduced rather than removed: disclose the existence of the generated content "in an appropriate manner that does not hamper the display or enjoyment of the work".

The second case is text. Deployers of a system "that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated". The trigger is the purpose behind the publication, not the topic alone. A product launch caption is not published to inform the public on a matter of public interest. A post from a bank or a clinic explaining a policy change to the people it affects sits close to the line, which is why teams in regulated industries decide this deliberately rather than by default.

Does human review remove the disclosure duty?

Only for text, and only on stated conditions. The carve-out in Article 50(4) reads: the text disclosure obligation "shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content".

Three limits on that sentence get lost constantly:

• It attaches to the text limb of Article 50(4) and does not touch the deepfake limb. No amount of human review makes an AI-generated video of a real person exempt.
• It requires both halves. A review process is not enough on its own. A named natural or legal person has to hold editorial responsibility for the publication.
• It does not reach Article 50(2). The provider's marking duty is unaffected by what any human downstream did.

"A human checked it" is not the test. The test is an editorial process with someone accountable for what went out, a structure a newsroom already has and most social calendars do not.

The text carve-out holds only if every link survives

AI-generated text

Human review or editorial control

Named person holds editorial responsibility

Article 50(4) disclosure not required

Who does Article 50 bind, and where?

Article 50 is EU law. Article 2(1)(b) applies the Regulation to "deployers of AI systems that have their place of establishment or are located within the Union". Article 2(1)(c) extends it to providers and deployers established in a third country "where the output produced by the AI system is used in the Union".

A US agency running an EU client's account is in scope through 2(1)(c), because the output is used in the Union. Article 3(4) also excludes an AI system "used in the course of a personal non-professional activity", so a private account posting an AI meme is not a deployer. A brand account is.

Article 50(5) governs delivery. The information "shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" and "shall conform to the applicable accessibility requirements". A label at the start of the video, not buried at the end of the caption.

What are the penalties?

Article 99(4)(g) puts "transparency obligations for providers and deployers pursuant to Article 50" in the tier carrying administrative fines "of up to EUR 15 000 000 or, if the offender is an undertaking, up to 3 % of its total worldwide annual turnover for the preceding financial year, whichever is higher". Article 99(6) reverses that for SMEs including start-ups, capping the fine at whichever figure is lower.

What is still unsettled?

Two things, both worth knowing before you build a process on guidance.

Article 50(7) tasks the Commission with encouraging codes of practice on "the detection, marking and labelling of artificially generated or manipulated content", and lets it adopt an implementing act specifying common rules if it judges a code inadequate. The code is voluntary. The AI Act Explorer's own guide to Article 50, dated 14 May 2026, describes both the final Code of Practice and the final Commission Guidelines as expected rather than published, and the site gives no confirmed adoption date for either. Elements under discussion there, including a standardised "AI" visual label localised per language and a split between "fully AI-generated" and "AI-assisted" content, are proposals, not rules.

The same guide states that "the AI Omnibus provisional agreement of May 2026 grants generative AI systems already on the market before that date until 2 December 2026 to meet the machine-readable marking requirement under Article 50(2)". The site's own consolidated text of Article 113, which does carry amendment markers for the Article 5 and Article 6 dates, shows no such transitional for Article 50. Either way, that transitional runs to Article 50(2) and therefore to providers. Deployers got no grace period on Article 50(4).

This page is not legal advice.

What this means for a social calendar

If you write captions with an AI caption writer or produce visuals with an AI social media image generator, you are a deployer and Article 50(4) is your paragraph. The work is deciding, per post, whether the asset is a deepfake under Article 3(60) and whether the text informs the public on a matter of public interest, then recording that decision where you can retrieve it. A social media content calendar is where that record lives, and agencies need it per client, because Article 2(1)(c) turns on audience.

Frequently Asked Questions

Does Article 50 require me to label every AI-written caption?

No. Article 50(4) reaches AI-generated text only where it is "published with the purpose of informing the public on matters of public interest". Marketing copy falls outside that limb. The marking duty in Article 50(2) sits on the provider, not on you.

Is an AI-generated image automatically a deepfake?

No. Article 3(60) defines a deep fake as content that "resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful". An invented product scene resembling nothing real fails that test. A synthetic photo of a named executive at a real venue passes it.

My model provider watermarks its outputs. Am I covered?

No. The watermark discharges Article 50(2), machine-readable marking read by detection tools. Your duty under Article 50(4) is a disclosure to the person looking at the post, delivered under Article 50(5) "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure".

Does Article 50 apply to a US company?

It applies through Article 2(1)(c) to providers and deployers established in a third country "where the output produced by the AI system is used in the Union". A US agency posting for an EU client is in scope; a US brand with no EU audience is not. Article 50 is EU law and does not bind publication that never reaches the Union.

Does an editor approving the post remove the disclosure duty?

Only for the text limb, and only where the content "has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication". Both halves are required. The carve-out never applies to deepfakes.

What happens if we get it wrong?

Article 99(4)(g) subjects Article 50 breaches to fines "of up to EUR 15 000 000 or, if the offender is an undertaking, up to 3 % of its total worldwide annual turnover for the preceding financial year, whichever is higher". Under Article 99(6), SMEs and start-ups are capped at the lower figure.

What is Chapter IV of the EU AI Act?

Chapter IV of the EU AI Act is titled "Transparency Obligations for Providers and Deployers of Certain AI Systems," and it holds all seven paragraphs of Article 50. It reaches any AI system used in one of four named situations, not the narrower set of high-risk uses Chapter III defines. A brand that runs no high-risk AI still falls inside it.

Does Article 50 have a grace period for AI content marking?

Only for providers, and only if a disputed transitional provision holds up. The AI Act Explorer's May 2026 guide says the AI Omnibus provisional agreement gives generative AI systems already on the market before that date until 2 December 2026 to meet the Article 50(2) marking rule, but the site's own consolidated text of Article 113 shows no such transitional. Either way, that grace period would run to providers under Article 50(2), not to deployers. Deployers get no grace period on the Article 50(4) disclosure duty.

How and when does an AI disclosure have to be shown?

Article 50(5) requires the disclosure to reach the person seeing the content in a clear and distinguishable manner, at the latest by the time of first interaction or exposure. It also has to meet applicable accessibility requirements. In practice that means a label at the start of a video, not text buried at the end of a caption.

Are small businesses fined the same as large companies under Article 50?

Article 99(4)(g) sets the standard Article 50 fine at up to EUR 15 000 000 or 3 % of worldwide annual turnover, whichever is higher. Article 99(6) flips that formula for SMEs and start-ups, capping their fine at whichever of those two figures is lower.