The reading that has spread fastest since August — that anything touched by AI now needs a visible “made with AI” notice — is wrong. Acting on it costs you more than it protects you, and it draws attention away from the two duties that do bite.
Article 50 of the AI Act contains four separate obligations. Two fall on providers, two on deployers. Most companies are deployers: you use an AI system someone else built. Sorting out what you owe starts with knowing which of the four you are inside.
The date that has passed, and the one still ahead
Article 50 has applied since 2 August 2026. Systems placed on the market before that date have until 2 December 2026 to meet the machine-readable marking requirement in Article 50(2). Content generated before 2 August 2026 does not have to be labelled retroactively, although the Commission encourages it.
The Commission’s Code of Practice on transparency of AI-generated content was published on 31 July 2026 and has been assessed as adequate by both the Commission and the AI Board. It has two sections that can be signed separately: one for providers covering marking, one for deployers covering deepfakes and public-interest text. Around 190 organisations had signed by the end of July, roughly 150 of them as deployers, and it remains open for signature. Two implementation task forces begin work in September 2026.
Which of the four duties is yours?
Article 50(1) binds providers: a system that interacts directly with people must tell them they are dealing with an AI system, unless that is obvious to a reasonably well-informed person. The notice must appear from the start of the first interaction.
Article 50(2) binds providers of generative systems: synthetic audio, image, video and text must be marked in machine-readable form. Article 50(3) binds deployers of emotion recognition and biometric categorisation systems, who must inform the people exposed to them. Article 50(4) and 50(5) bind deployers: deepfakes must be labelled, and so must AI-generated text published to inform the public on matters of public interest.
If you build or configure a customer-facing assistant, the first duty is yours. If you buy one, it is your vendor’s — but the notice appears on your website, under your name, and it is your visitors who are affected if it is missing.
What a provider has to mark, and how
Marking under Article 50(2) is technical provenance data — watermarking and embedded signals — not something a reader sees. It is routinely confused with labelling, and the confusion matters, because it sends companies looking for a visible notice they do not owe.
For a deployer it converts into a procurement question rather than a content question: do the tools your teams use mark their output, and can the vendor show you how?
Several categories sit outside the duty: short sequences of numbers, symbols or letters; source code; output exchanged between machines with no human exposure; closed-loop industrial and product development output; and standard assistive editing functions. A grammar tool that tidies a sentence is not generating synthetic content.
When does a deepfake need a label?
Article 50(4) covers image, audio or video that resembles real people, objects, places or events and would falsely appear authentic. The label must be clear and appear at first exposure at the latest. Artistic, satirical and fictional works get a lighter form: disclosure in a manner that does not hamper the display or enjoyment of the work.
This is the duty that reaches marketing most directly. A synthetic presenter who reads as a real person, a generated cityscape presented as an actual location, a voice built to resemble a named individual — all inside. A plainly stylised illustration is not a deepfake, because it does not purport to be authentic in the first place.
The human review that removes the duty
Article 50(5) covers AI-generated or manipulated text published to inform the public on matters of public interest: politics, administration, justice, fundamental rights, security, health, environment, consumer safety, and economic, financial, scientific or cultural developments.
Text that has undergone genuine human review or editorial control needs no label. Review means deliberate examination by someone with the knowledge to exercise professional judgment. Editorial control means the authority to approve, alter or reject content on substantive grounds. A spell check does not qualify.
Ordinary commercial copy is rarely published to inform the public on a matter of public interest to begin with. Where content does touch those subjects, the exemption is available to any organisation with a named person approving what goes out. What matters is that the review is real and that you can show it happened.
Practical example: the Malmö software company
A company in Malmö sells a scheduling tool to Swedish and German customers. It runs a support chat built on a third-party assistant, publishes a monthly article on workplace regulation drafted with a general AI tool and edited by its head of communications, and uses AI-generated stock imagery in campaigns.
The support chat falls under Article 50(1), so the company has to confirm that the vendor’s notice appears at the start of the first interaction. The articles fall potentially under Article 50(5), but the editor’s review exempts them, provided the company can show the review took place. The imagery stays outside Article 50(4) as long as it does not present invented people or places as real. What remains is the vendor question under Article 50(2) and a record of who edits the articles. Neither takes long.
Vanliga misstag
Treating marking and labelling as the same duty, and adding visible notices where machine-readable marking was what the article required.
Assuming a support chat is obviously an AI system because everyone internally knows that it is.
Relying on the human-review exemption without any record of who reviewed what, and when.
Reading Article 50 as a duty on the AI vendor alone, when paragraphs 3, 4 and 5 sit with the deployer.
Labelling every AI-assisted sentence, which dilutes the notices that carry real information.
Rekommenderade åtgärder
Ask every generative AI vendor how their output is marked, and get the answer in writing.
Check whether any system in service before 2 August 2026 will meet the marking requirement by 2 December 2026.
Confirm that your customer-facing assistants identify themselves at the start of the first interaction.
Name the person who reviews published content on public-interest subjects, and record that the review happened.
Decide whether to sign the Code of Practice, including its deployer section.
Vanliga frågor
Does Article 50 apply to AI we only use internally?
The transparency duties are triggered by the exposure of natural persons. A system that only talks to other systems, with no person on the other end, falls outside the article. Internal use that puts employees in front of an AI system, or that produces content which is then published, can still engage it.
We use a general AI assistant to draft marketing copy. Must we label it?
Normally not. Article 50(5) reaches text published to inform the public on matters of public interest, which most commercial copy is not, and genuine human review exempts the text even where it does. The marking duty under Article 50(2) is a separate matter and sits with the provider of the tool.
Which authority supervises this in Sweden?
On 12 June 2026 the government designated five national competent authorities under the AI Act: Post- och telestyrelsen, Integritetsskyddsmyndigheten, Finansinspektionen, Läkemedelsverket and Swedac. The assignment runs until 31 December 2026 and permanent legislation is still being prepared. Penalties for breaching Article 50 reach EUR 15 million or 3% of total worldwide annual turnover, with proportionality applied to SMEs and small mid-cap companies.
Slutsats
Most of what a marketing function produces with AI needs no visible label. What it does need is an accurate answer from every vendor about marking, and a documented human review where public-interest content is published. Both are quick to put in place now and awkward to reconstruct once an authority is asking. The narrow reading of Article 50 is also the correct one, and it leaves you with a shorter list than the panic suggests.
If you want a clear picture of which of your systems and content sit inside Article 50, and what to prioritise, we will go through it with you. The first hour is on us.
