AI Product Photos in Ads: When the EU Calls Them Deep Fakes

On 12 November the European Commission's guidelines on Article 50 of the AI Act appeared in the Official Journal as C/2026/6005. The text itself is not new: the Commission first released it on 20 July, two weeks before Article 50 started to apply. What the formal publication does is put the Commission's reading of the rules in a place lawyers, regulators and procurement teams will cite for years.

We covered the chatbot side of Article 50 earlier this month. This post is about a narrower question that affects far more marketing teams: when does an AI-edited product photo in an ad become a deep fake that needs a visible label?

The short answer from the guidelines: when the image can make people believe the product looks, works or performs differently than it really does. Using AI is not the trigger. Changing the product is.

Where the Commission draws the line

The AI Act defines a deep fake as AI-generated or manipulated image, audio or video that resembles existing people, objects, places or events and would falsely appear authentic. Deployers, meaning the companies that use an AI system under their own authority, must disclose such content under Article 50(4).

Most people read "deep fake" as fake politicians and cloned celebrity voices. The guidelines go further. Among their examples of a deep fake, as quoted in PPC Land's detailed analysis of the published text, is "an AI-generated image of a product in advertisement or packaging that can affect the audience's perception and mislead as to the actual product appearance, characteristics or use", for example by making the product look "more appealing or with improved quality than in real life".

On the other side of the line, the same text says AI colour correction, background extensions, background replacements for clearly aesthetic purposes, arrangements of existing products and re-scaling are likely to have only a minor effect on whether people perceive the image as authentic. A real car placed in an AI-generated landscape is not a deep fake as long as the ad does not mislead about the car itself. An obviously fantastical scene, such as mice arguing about cheese in a cheese commercial, is not one either.

Three further points matter for anyone running campaigns:

• Intent does not count. The assessment is objective. A team that "only wanted a nicer picture" is judged on the result, not on its motive.
• The audience is wide. The guidelines ask deployers to think about the realistic audience, including children and older people, rather than a well-informed average viewer.
• A watermark is not a label. Deep fake labels must be perceivable by people without special tools, so a deployer cannot rely on the machine-readable mark the image generator embedded.

The guidelines are non-binding, and only the Court of Justice can give an authoritative interpretation. But national market surveillance authorities will start from this text, and so will most legal advice.

A sofa, two prompts

A hypothetical example makes the line easier to see. An online furniture shop has a good studio photo of a grey sofa and wants a campaign image for social media.

Prompt one: "Place this sofa in a bright Scandinavian living room with morning light." The sofa stays as it is. The room is invented, but nobody buying the sofa is misled about the sofa. On the guidelines' reading this is an aesthetic background change and does not need a deep fake label.

Prompt two: "Make the fabric look softer and fuller, remove the creases and make the grey a little warmer." Now the product in the image is not the product in the warehouse. A customer who receives a flatter, cooler grey sofa was shown something that does not exist. This is exactly the kind of image the Commission lists as a deep fake, and it needs a clear label at the moment people first see it.

There is a second problem with prompt two that a label does not solve. The guidelines state that labelling does not make lawful content that is unlawful under other rules, and they name misleading advertising explicitly. A small "AI-generated" tag in the corner does not protect an ad that misrepresents the product under consumer law. For product images, the honest fix is usually to not change the product rather than to label the change.

Who has to put the label on

Article 50(4) applies to deployers. The guidelines use an advertising company as their example: the agency remains the deployer even when freelancers or contractors operate the tools for it, and its designers and content creators are not separate deployers.

The client side gets a carve-out with a condition. A company that merely commissions an agency, without deciding whether and how AI is used in production, is not a deployer. Once the brand chooses the tool, writes the prompts or sets the parameters, it can become one.

If your team makes product images in-house with a generative tool, there is no agency to carry the duty. Your company is the deployer.

Geography does not help much either. According to the guidelines, a company outside the EU that publishes deep fakes on the open internet, foreseeing that people in the Union will see them, is covered.

What 2 December changes and what it does not

You will see 2 December 2026 mentioned a lot in the coming weeks. It is the end of a grace period, but a narrow one. As the Commission's FAQ explains, generative AI systems placed on the market before 2 August 2026 have until 2 December to meet the machine-readable marking and detection duty under Article 50(2).

That is an obligation for the providers of image and video generators, not for the shop using them. The deployer's duty to label deep fakes has applied since 2 August. Content generated before 2 August does not have to be labelled retroactively, although the Commission encourages it where possible.

Fines for breaching Article 50 can reach €15 million or 3% of worldwide annual turnover, whichever is higher; for SMEs and start-ups the lower amount applies. In Bulgaria the practical question of who will enforce this is still open: the country has not yet adopted its institutional framework for the AI Act, and after two failed attempts in parliament a third bill is expected, according to Ivo Emanuilov of the GATE Institute speaking to BNR. A missing national regulator delays enforcement. It does not suspend the obligation.

A rule a marketing team can actually follow

My reading of the guidelines is that companies do not need a long AI policy for product images. They need one rule that a designer can apply at 6 pm on a Friday: AI may change everything around the product, but not the product itself.

In practice that comes down to a few habits:

• Keep the original product photo next to every AI-edited version, so anyone can compare what changed.
• If the product's shape, colour, size, texture or function was altered, either revert the change or do not use the image in advertising or on packaging.
• For the rare case where an altered depiction is deliberate and not misleading, agree in advance on a visible label and where it appears, including on short video where people may not see the first seconds.

The middle one does most of the work. It also keeps the brand on the right side of consumer law, which applies whether or not anyone mentions the AI Act. If you want to turn this into a written rule alongside your other AI usage policies, our AI governance work covers exactly this kind of policy.