What Article 50 of the AI Act says about content

Article 50 of Regulation (EU) 2024/1689 splits the work between two roles. The provider develops the AI tool and places it on the market under its own name. The deployer uses it under its own authority, outside purely personal, non-professional use (Article 3(4)). A small business creating images with AI for its website is a deployer.

Paragraph 2 requires providers to make sure the text, images, audio or video their tool generates are marked in a machine-readable format and detectable as artificial. Paragraph 4 requires the people publishing to disclose two things, deepfakes and text that informs the public on matters of public interest.

Article 113 makes all of this apply from 2 August 2026. The European Commission guidelines of 5 November 2026 say any activity through which someone earns money on a regular basis counts as professional (point 19). A freelancer creating AI images for their business is covered too.

What needs a label, content type by content type

It depends on the type of content and on whether it could make someone believe it is real. The table sums up Article 50 and the examples in the Commission guidelines, which explain how the Commission reads it. Use it as a first filter. If your case isn't listed, compare it with the closest example.

Content created or edited with AI and what the AI Act requires, based on Article 50 and guidelines C(2026) 7702
ContentWhat the rules sayWhat you need to do
Product photo colour-corrected, resized or given a new background for looks

Minor edit that usually doesn't make it a deepfake (guidelines, point 116)

No mandatory label if it doesn't mislead about the product

AI-generated product image that makes it look different or better

Deepfake example in the guidelines (section 6.1.1)

Visible label since 2 August 2026

Realistic synthetic person or spokesperson in an ad or video

Deepfake example in the guidelines (section 6.1.1)

Visible label since 2 August 2026

Photos of an empty flat furnished with AI

EU example for the partially modified content icon

Visible label

Clearly fantastical scene, such as mice having an argument

Not a deepfake according to the guidelines (section 6.1.1)

No mandatory label

Product description or ad copy written with AI

Outside Article 50(4) unless it makes health, safety or sustainability claims (guidelines, section 6.2.1)

No mandatory label

Article on health, politics or public services written with AI

Text on matters of public interest (Article 50(4))

Label, or human review with someone holding editorial responsibility

Invisible mark inside the file

The tool provider's obligation (Article 50(2))

Ask the provider whether it already does this

If I write my blog with ChatGPT, do I have to say it's AI?

Only if the text informs the public on a matter of public interest, such as health, politics or a public service, and nobody has properly reviewed it. If a person checks the facts, edits it substantively and someone in the business is accountable for it, Article 50(4) doesn't require a label. Ad copy and product text are outside its scope, according to the guidelines.

What actually changes on 2 December 2026

2 December 2026 only affects providers. Article 111(4), added by Regulation (EU) 2026/1744, gives systems that generate audio, images, video or text and were placed on the market before 2 August 2026 until that day to comply with Article 50(2), the machine-readable marking rule.

There is no such extension for visible labels on deepfakes and public-interest text. The guidelines say Article 50 has applied to all systems since 2 August 2026, and tools that both chat and generate content only get extra time for marking (point 153). Telling people they are talking to an AI was already mandatory in August.

Nor do you need to label what was made earlier. According to the guidelines, content and deepfakes generated before 2 August 2026 don't need retroactive marking or labelling (point 154). Public-interest text generated before that date but published after it does need a label.

For a small business, 2 December is mostly a reminder to ask your tools whether they already mark what they generate. Google, for example, says it embeds invisible SynthID watermarks and C2PA metadata in images and videos generated with Google Ads tools.

Do AI-generated stock images need a label?

It depends on what they show. If they look like real photos of people, places or events and someone could take them as authentic, they fit the Article 3(60) definition of a deep fake and whoever publishes them must disclose it. If they are clearly illustrations or impossible scenes, the guidelines give similar examples that are not deepfakes.

Do AI images on your website or social media need a label?

Only if they are deepfakes. 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. The guidelines add that no intent to deceive is needed (point 114).

The guidelines give marketing examples (section 6.1.1). An AI-generated product image that makes the product look different or better than it is counts as a deepfake, as does a synthetic influencer testing a product or a realistic avatar of the CEO thanking staff. A real car shown against a generated background doesn't, as long as it doesn't mislead about the car.

Everyday edits usually don't count. According to the guidelines, colour correction, removing a passer-by or swapping a background for aesthetic reasons tend to have little effect on how people perceive the content, though it depends on context (point 116). An empty flat furnished with AI, by contrast, is the example the EU uses for its partially modified content icon.

If the product has to look exactly right, the safest route is to start from the real thing, with photos or product renders built from its true dimensions and materials. If you're unsure whether an image could pass as real, add the label.

Want us to look at the AI images and ads you already have online? Leave your details and we'll tell you what we would check first.

Let's check

What about AI-written text that a person reviews?

If a person genuinely reviews it and someone is accountable for it, no label is needed. Article 50(4) only requires disclosure for AI-generated or manipulated text published to inform the public on matters of public interest. It makes an exception where the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility.

Most business text falls outside it. The guidelines say ad copy and product descriptions are not covered, provided they make no claims about health, consumer safety or sustainability. Covered examples (section 6.2.1) include an article on a lifestyle website about diets and disease, or investor information published by a listed company.

The review has to be substantive and done by someone who knows the subject. The guidelines say fact-checking is the minimum, while a spell-check, an editorial policy nobody applies or a quick sign-off don't count (points 134 and 135). If AI changes the text again after the review, the exception no longer applies (point 136).

A clinic blogging about a medical condition or a town council publishing notices with AI is much closer to the line. They should decide who reviews each text, put that in writing and, where there is no review, put the label at the start of the text.

Does a humorous ad made with AI need a label too?

If it is a deepfake, yes, but the requirement is lighter. When it is part of an evidently artistic, creative, satirical or fictional work, disclosure in a way that doesn't spoil the work is enough (Article 50(4)). The guidelines say ads only qualify in specific cases (point 122), and a video of synthetic consumers trying out a product to sell it does not.

What a label should look like and the EU icons

The label must be clear and distinguishable and reach people at the latest the first time they see the content (Article 50(5)). The guidelines say it isn't enough to put it in the terms of use, in a manual or deep in a menu, or to rely on the invisible mark in the file (points 117 and 142).

In a video, the label goes at the start, and again later if some viewers are likely to miss the beginning (point 143). On social media you can use the labelling tool that large platforms offer, but the guidelines say the responsibility stays with you (point 126).

The EU has published free icons for labelling content, in three versions (basic, fully AI-generated and partially AI-modified) and four colour variants. They are optional, work better with a short text label next to them and, according to the Commission, should be embedded in the content so they stay visible when it is shared or downloaded. Where you can, add alt text saying it was made with AI.

Fines for breaching Article 50 can reach EUR 15 million or 3% of total worldwide annual turnover for the previous financial year, whichever is higher (Article 99(4)). For SMEs, including start-ups, the lower of the two applies (Article 99(6)). If you also run a chatbot, the chatbot notice is a separate obligation under Article 50(1).

Is the Commission's code of practice mandatory?

No. The Commission published it on 10 June 2026 and it is voluntary, but Article 50 itself is binding. Signatories can rely on it to show they comply. Non-signatories have to show compliance by other means, and the guidelines say they may face more requests for information (point 148). By the end of July about 190 organisations had signed it.

What to check this week on your website, social media and ads

You can do this check in an afternoon and be ready before 2 December. If your team uses AI applied to marketing every week, keep the list and run it again whenever you change tools.

  • List the AI tools your team uses for images, video, voice and text.
  • Ask each provider whether it marks what it generates and since when, and keep the answer.
  • Go through the images and videos published since 2 August 2026 and set aside any that could look real.
  • Add a label or the EU icon to every deepfake, on the image and in the accompanying text.
  • Decide who reviews public-interest text, how they check the facts and who is accountable, and write it down.
  • In Google Ads, mark AI-made assets with the label setting and check how the ad looks.
  • Record the date of the check and who did it.

Add crida studio as a preferred source on Google