An agency retouches a website photo and swaps the person in it for someone generated with AI. The marketing team plans a campaign with a model who does not exist. Another client asks a tool to “improve” a product image. In every case we get the same question: when is labelling AI-generated images mandatory?
Since 2 August 2026 the answer turns on one idea. If the image looks real and is not, you must disclose it, and the metadata added by the tool is not enough. If it is obviously fantasy, or the edit is minor, no label is needed. Here is where the line falls under the European Commission’s Guidelines and the Code of Practice, with examples and an anonymised real case.
In this article we will discuss...
What Article 50 of the AI Act says and when it applies
Article 50(4) of Regulation (EU) 2024/1689 applies to the deployer, meaning whoever uses the AI tool in a professional capacity. The deployer must disclose that image, audio or video content constituting a deep fake has been artificially generated or manipulated.
Under Article 3(60), a deep fake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events. It must also be capable of falsely appearing to a person to be authentic or truthful. So there is no need for a celebrity, and no need for an intention to deceive.
Three points to be clear about
- The disclosure must be clear and distinguishable, given at the latest at the first exposure, and meet accessibility requirements (Article 50(5)).
- Technical marking is not enough. Article 50(2) requires the tool provider to mark outputs in a machine-readable format. However, the Commission states that the deployer cannot rely on that marking, because nobody can see it without specific tools.
- Dates. Article 50 applies from 2 August 2026. The AI Omnibus Regulation gives providers of tools already on the market until 2 December 2026, and only for technical marking. The company’s visible label gets no extension.
Is your company already using AI to create images, video or voices? We review your uses and give you a labelling protocol ready for marketing and for your agency. Artificial intelligence lawyers for companies.
Why this obligation exists
The Commission Guidelines refer to recitals 132 to 136 of the AI Act. According to them, the aim is to reduce the risks of impersonation, deception, disinformation, manipulation and fraud, and to protect trust in the information ecosystem.
This matters. The rule is not only about consumer protection, which misleading advertising law already covers. The Guidelines expressly say that the “falsely appearing authentic” criterion is assessed independently of the concept of deception in the Unfair Commercial Practices Directive. In other words, an image may not be misleading advertising and still need a label.
The “does it look real?” test: the Commission’s four criteria
The Guidelines break the definition into four cumulative criteria:
- Appreciable resemblance. It need not be identical; a high level of similarity to the subject is enough.
- Persons, objects, places, entities or events. This includes realistic AI-generated people, buildings, consumer goods, animals and scenes such as the provision of a service.
- Existing. Here is the surprise: it covers what exists and also what could plausibly exist. A realistic invented person is covered. Dragons, humans flying unaided or elephants driving cars are not.
- Capable of falsely appearing authentic. This is assessed as a whole, by context and foreseeable audience. Moreover, the test is not an “average” person; children, older people and people with lower digital literacy also count.
Note what is not a deep fake. The Guidelines list lighting adjustments, colour correction and noise reduction as minor edits. The same goes for removing a passer-by from the background or background changes for clearly aesthetic purposes. Even so, context matters: in a news photo, editing the background beyond standard practice can count.
Examples: which images need a label and which do not
A label is required
The first is a frequent case in our practice; the rest are examples from the Guidelines themselves:
- A real photo of the business where a person has been replaced by an AI-generated one.
- A realistic synthetic model or influencer trying or promoting a product.
- A video with a realistic avatar of the CEO congratulating staff.
- An AI-generated product image that makes the product look different, more attractive or better quality than it really is.
- A teleshopping-style video with simulated people using the product. The Guidelines expressly exclude it from the lighter regime for creative works.
No label is needed
By contrast, the Guidelines do not treat these as deep fakes:
- An image of a sphinx flying over the Eiffel Tower.
- An advert with talking mice discussing the best cheese.
- A real product on an AI-generated background, provided the advert does not mislead about the product.
- Lighting, colour or noise adjustments that leave the essentials unchanged.
- Fictional video game environments or a cartoon of a historical event.
What about “creative” advertising?
Evidently artistic, creative, satirical or fictional works have a lighter regime: you must still disclose, but in a way that does not spoil the work. However, the Commission reads it strictly. An advert only benefits if its creative nature is evident to the public, and where informative and creative elements are mixed, the informative character prevails. In practice, most product campaigns will not qualify.
AI-generated or AI-modified: not the same thing
The Code of Practice on Transparency of AI-Generated Content, published on 10 June 2026, offers three EU icons. One is for fully AI-generated content (“AI + GENERATED”), another for partially modified real content (“AI + MODIFIED”) and a third, basic one with just “AI”. The Code is voluntary, but on 8 July 2026 the Commission concluded that it is adequate to demonstrate compliance.
The difference has consequences. If the photo is real and only a person has been changed, the right label is “modified”, not “generated”. Saying “generated” is not just inaccurate: it suggests that the premises, the product or the team are invented, which damages the brand.
Where and how to label AI-generated images
For signatories of the Code, the specifications are concrete:
- The main element is the capitalised acronym “AI”. Signatories are encouraged to add whether the content is generated or modified and, where possible, what has changed (for example, “a face has been altered”).
- Embedded in the image itself, unless equivalent alternatives exist, and visible at the latest at first exposure.
- In a place free of other overlays, for example the top right corner, and readable against any background.
- In video: at the beginning and, at a minimum, after each interruption.
- Accessible: high contrast, compatible with screen readers, with alternative text.
A notice hidden in the terms of use or in a menu does not comply. The Guidelines are clear: if people can easily miss it under normal conditions, it is not clear and distinguishable.
Do images from before 2 August 2026 need labelling?
Not retroactively. The Commission says that deep fakes generated or manipulated before that date do not need labelling, although it encourages doing so without disproportionate effort. For example, it does not expect companies to audit content databases or reprint packaging.
Our practical advice goes further. If the image is still being circulated, such as in a monthly newsletter or on the home page, labelling it costs little and avoids arguments about when it was edited. And, in our view, if it is manipulated with AI again after August, it is new content.
Who is responsible: company, agency and employees
The deployer is whoever uses AI under their authority, meaning whoever decides why and how it is used. Technical control is not required. Here is another point few people explain:
- The agency that generates or modifies the image is a deployer.
- The client company is also a deployer if it decides on or approves the use of AI. By contrast, under the Guidelines, a company that merely commissions an advert, without deciding or controlling whether the agency uses AI, is not a deployer.
- Employees of the agency or the company (designers, content creators) are not separate deployers.
That is why we recommend putting it in writing in the agency contract. It should state who decides to use AI, who applies the label and how the label is kept when the image is reused. The Guidelines themselves cite contractual terms as a reasonable measure.
Fines and competent authority
Breaching Article 50 can lead to fines of up to €15 million or 3% of worldwide turnover, whichever is higher. For SMEs and start-ups, the lower of the two applies. In addition, adherence to the Code can be taken into account as a mitigating factor.
There is an important point in Spain. The AI Act applies directly. However, the draft Organic Law on the proper use and governance of AI, which was meant to set out the authorities and sanctioning procedure, lapsed when Parliament was dissolved in October 2026. The Spanish AI agency (AESIA) has published explanatory materials on Article 50, but no criteria of its own. That gap is not a licence: the obligation exists, and an unlabelled image can also lead to a complaint for misleading advertising or under data protection law.
Case study: a real photo with a person replaced by AI
An optician had photos from a real shoot in its consulting room on its website and in its newsletter. A former employee appeared playing a customer and asked to be removed. The agency replaced her with an AI-generated person, with the optician’s approval.
- Data protection: solved, because the former employee no longer appears.
- Article 50: the photo still looks authentic. The consulting room is real, the logo is visible and the person is realistic. Moreover, an essential element was changed, not a background detail. It is a deep fake.
- Who is responsible: the agency, which used the AI, and the optician, which decided on the change and publishes the image.
- Correct label: “Image modified with AI”, clearly visible and embedded in the image, with the “AI” icon if following the Code.
The client believed that, because the photo is real, nothing was needed. It is the most common mistake: the more real the base, the more necessary the label.
If someone has to be removed from your photos, such as a former employee, see also what to do with a former employee’s email account and when a former employee asks for their data.
Checklist for companies
- Map where you use AI for images, video or voice: website, social media, catalogues, newsletter, packaging, internal training.
- Apply the test: does it resemble something real or plausible, and could it pass as authentic to your audience?
- Distinguish between generated, modified and minor edits.
- Define the label and its position, embedded in the image and visible from first exposure.
- Agree with your agency who decides, who labels and who keeps the label.
- Train your marketing and communications team. The Code calls for proportionate awareness measures.
- Document your approach and examples. If you do not sign the Code, the Commission expects you to explain how you comply.
- Also review data protection and image rights for anyone shown or imitated.
How we can help
At Auratech Legal Solutions we review how your company uses AI and give you a labelling protocol for your website, social media and campaigns. We also draft the clause for your agency contract and train your marketing team in AI literacy. If you already have published images and are unsure whether to label them, tell us about your case.
More on our pages on artificial intelligence lawyers and new technology and advertising law.
Frequently asked questions
Do all AI-generated images need a label?
No. Only those that resemble real or plausible persons, objects, places or events and could pass as authentic. An obviously fantastical illustration or a minor lighting or colour edit needs no label.
Is the tool’s metadata or watermark enough?
No. That technical marking is applied by the provider and cannot be seen without specific tools. The company publishing the image must add a visible, understandable disclosure.
Does an AI-invented person count as a deep fake?
Yes, if realistic. The Commission includes what could plausibly exist, so a synthetic model who looks real is covered.
Which label should I use if I only changed part of a real photo?
The correct disclosure is that it was modified with AI, not generated. The Code provides a specific icon for partially modified content.
Do images published before 2 August 2026 need a label?
Not retroactively, although the Commission recommends it. If they are still being circulated, labelling them costs little and avoids disputes.
If the agency made the image, is the company also responsible?
Yes, if it decided on or approved the use of AI. Only a company that commissions an advert without deciding or controlling whether the agency uses AI falls outside.
What is the fine for not labelling?
Up to €15 million or 3% of worldwide turnover, whichever is higher; for SMEs, the lower. In Spain, the law that was to set out authorities and procedure lapsed in October 2026.
Sources: Regulation (EU) 2024/1689, Articles 3(60), 50 and 99; Commission Guidelines on Article 50 (20 July 2026); Code of Practice on Transparency of AI-Generated Content (10 June 2026); Commission Q&A.
This article provides general information and is not a substitute for legal advice on a specific case.



