Do you need to label AI-generated content? Article 50 of the EU AI Act for small businesses

As of 29 October 2026. This is not legal advice. The article describes the EU AI Act and the European Commission’s guidelines. For a specific case, speak to a lawyer.

The date 2 December 2026 is easy to misread as “the day businesses must start labelling AI content”. That deadline is for the companies that make AI tools. The duties of businesses that use those tools have applied since 2 August 2026, and they cover only some kinds of content. Here is how it breaks down.

Provider or deployer: everything depends on your role

The AI Act splits duties by role.

• A provider (Article 3(3)) develops an AI system, or has one developed, and places it on the market under its own name or trademark. Think of a company selling an image generator or a writing assistant.
• A deployer (Article 3(4)) uses an AI system under its authority, other than for purely personal use. Think of a shop generating images for adverts, or a firm drafting blog posts with AI.

Most small businesses are deployers. A business can become a provider if it builds its own system, or modifies someone else’s and offers it under its own brand. That can happen with a chatbot built to order.

The Commission’s guidelines add two practical points. Employees and contractors working under a business’s instructions are not separate deployers: the business is responsible. And a business that simply commissions an advert from an agency, without deciding whether or how the agency uses AI, is not a deployer. The duty then sits with the agency.

What the 2 December 2026 deadline covers

Article 50(2) requires providers of systems that generate audio, images, video or text to mark the outputs in a machine-readable format so they can be detected as artificially generated. These are technical markers, for example in metadata or in the file itself, and people cannot see them.

Regulation (EU) 2026/1744 added Article 111(4). Providers whose systems were placed on the market before 2 August 2026 have until 2 December 2026. A business that only uses such a tool does not have to add those markers itself.

The provider duty does not cover tools that only assist with standard editing or do not substantially change the input or its meaning.

What a business must label: deepfakes

Article 50(4), first subparagraph: a deployer of an AI system that generates or manipulates image, audio or video content constituting a deepfake must disclose that the content has been artificially generated or manipulated.

Under Article 3(60), a deepfake is 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. The guidelines explain that resembling something that exists, or could plausibly exist, is enough. A dragon or a person flying unaided is not realistic.

Examples of deepfakes from the guidelines that matter to businesses:

• a video with a realistic avatar of the company’s CEO thanking staff for the year’s results;
• a video with an AI-generated likeness of a celebrity in an advert;
• a product image in an advert or on packaging that misleads about how the product looks, what it does or how it is used, for example by making it look better than in real life.

Examples that are not deepfakes:

• a real product against an AI-generated background, as long as it does not mislead about the product;
• colour correction, removing a passer-by from the background, noise reduction, resizing;
• an obviously fictional advert, such as mice arguing about cheese.

Whether something is a deepfake depends on the whole picture: how close the resemblance is, the message, where it is published and who sees it. The guidelines say to take into account children and older people if they may see the content.

For evidently artistic, satirical or fictional works, it is enough to disclose in an appropriate way that does not hamper enjoyment of the work.

What a business must label: text on matters of public interest

Article 50(4), second subparagraph, covers AI-generated or manipulated text that is published with the purpose of informing the public on matters of public interest. The guidelines break this into three conditions:

1. Published: available to an indeterminate, fairly large audience. An internal intranet or correspondence with a client is not published.
2. Informing: the text communicates knowledge, opinions or facts. A short slogan does not.
3. Public interest: among others, politics, public administration and services, justice, public health, consumer safety, the environment, and economic, financial or scientific developments that may be the subject of public debate.

According to the guidelines, these fall outside the scope:

• advertising copy and product descriptions, unless they contain claims about health, consumer safety or sustainability;
• advice written by a consultant for a single client.

In scope, for example: an article on a lifestyle website comparing the effect of diets on a disease, or investor information published on a listed company’s website. A business guide to new regulations or to health topics may also fall within it. That is a case-by-case assessment.

The exception: human review and editorial responsibility

Text does not need a label if both conditions are met: it has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for the publication. The guidelines set a high bar:

• review means a substantive check by someone with relevant knowledge; fact-checking is the minimum;
• spell-checking, the mere existence of an editorial policy, automated review or cursory approval do not count;
• if AI substantially changes the text again after sign-off, the exception no longer applies;
• the identity and contact details of the person or function responsible for the publication should be easy to find, for example in the legal information on the website.

The exception covers text only. A deepfake must be labelled even if it went through careful editing. A business can also take the simpler route and label all AI-assisted texts, without documenting a review process.

How to label so that it counts

Article 50(5) requires the information to be given in a clear and distinguishable manner, at the latest at the first interaction or exposure, and in line with accessibility requirements. The guidelines and the code of practice translate this into practice:

1. Visible or audible to people. Metadata is not enough.
2. At the start. For text: above it or next to the headline. For video: at the start and, where possible, during it, at least after breaks. For audio only: a short spoken notice at the start.
3. Not in the terms and conditions. A notice hidden in a manual, a menu or terms nobody reads is not clear.
4. On each piece of content. A general “we use AI on this website” is not enough.
5. Accessible. Real text rather than only an image, sufficient contrast, alt text or an ARIA label, readable by screen readers.
6. Plain language. No jargon and no abbreviations other than “AI”.

The European Commission has published EU icons for labelling AI content: a basic icon, “fully AI-generated” and “partially AI-modified”, in black and white versions. Anyone can use them free of charge and without attribution. User testing showed that the icon works better with a short text label. Using the icon does not, on its own, make you compliant.

Code of practice. A code of practice on transparency of AI-generated content was published on 10 June 2026. The Commission and the AI Board have confirmed it is an adequate tool for demonstrating compliance. Section 2 is for deployers. Signing is voluntary. By the end of July 2026, about 190 companies and organisations had signed it. A business that does not sign must, if asked, show by other means that its labelling is adequate.

What about content from before 2 August 2026?

Under the guidelines, content generated and published before 2 August 2026 does not need to be labelled retroactively. But text prepared earlier and published on or after that date must be labelled. For older deepfakes, labelling is encouraged, without disproportionate effort such as auditing whole archives.

Ireland and the UK

Ireland is an EU member state, so the AI Act applies directly. The Regulation of Artificial Intelligence Act 2026, signed on 21 July 2026, gives further effect to it in Irish law and provides for the AI Office of Ireland (Oifig IS na hÉireann). Supervision is shared between sectoral regulators. Under S.I. No. 405 of 2026, Coimisiún na Meán supervises Article 50 where the provider or deployer is an intermediary or hosting service, a broadcaster, an audiovisual media service or a designated online service.

The UK. The AI Act was adopted in 2024, after the UK left the EU, so it is not part of UK law. It can still reach UK businesses. Under Article 2(1)(c), it applies to deployers outside the EU where the output of the AI system is used in the EU. The guidelines explain that this covers a deployer that expects its content to be distributed in the EU, for example by publishing deepfakes on the open internet. It does not cover content that reaches EU audiences through channels the deployer cannot foresee or control. In practice, a UK business with customers or audiences in Ireland or elsewhere in the EU should apply the same rules to that content.

Enforcement and fines

Breaching the Article 50 obligations can lead to a fine of up to EUR 15 million or 3% of worldwide annual turnover. For small and medium-sized enterprises, the lower of the two applies. Authorities can act on their own initiative or after a complaint, which anyone with grounds to suspect an infringement may lodge.

The AI Act does not displace other law. An advert with a product image that looks better than the real thing may also be assessed as misleading under consumer protection rules.

A checklist for your business

1. List where your business uses AI to create content: images, video, voice, website text, social posts, newsletters.
2. For each use, settle your role: are you using someone else’s tool, or offering your own under your brand?
3. Review images and video: do they show realistic people, products or events that viewers could take as real?
4. Check product photos: has AI changed how the product looks, what it does or how it is used?
5. Review texts: which are published and deal with matters of public interest, such as regulation, health or finance?
6. Decide: label those texts, or run a documented review with clear editorial responsibility.
7. Agree one label design and where it goes: above the text, in the corner of an image, at the start of a recording.
8. Check the label on a phone, with a screen reader and after sharing on social media.
9. Agree with your agency and contractors who decides on AI use and who labels.
10. Write these rules into your company AI policy and brief your team.

How we can help

We review AI tools and draft AI policies: we find out which AI tools your team uses and on what terms, then draft a policy template that fits your business. Your management or legal adviser signs off the final wording. We also run AI adoption projects: from a task review and a pilot in one team to training your people.

Tell us what your business uses AI for. You will get a free quote within 24 hours, including weekends and public holidays. The price is fixed once quoted.