When do you need to label an AI visual, and when not?
This article explains the deepfake test, the form of an AI disclosure, the EU icons and the labelling of AI ads in Google Ads and Meta Ads.
25. October 2026
12 min
IntermediateThis article explains the deepfake test, the form of an AI disclosure, the EU icons and the labelling of AI ads in Google Ads and Meta Ads.
25. October 2026
12 min
IntermediateLabelling an AI visual means giving visible notice that content was created or modified by artificial intelligence. The AI Act requires a label only for deepfake content. A deepfake is an AI visual that comes across as an authentic depiction of an existing person, object, place or event. Cropping, colour correction and retouching are generally edits that carry no labelling obligation. What decides the obligation is the impression the final visual creates, not the use of AI as such. Article 50 of the AI Act (Regulation (EU) 2024/1689) applies from 2 August 2026. This article explains the deepfake test, the form of an AI disclosure, the EU icons and the labelling of AI ads in Google Ads and Meta Ads.
No. Using AI to create an ad does not, in itself, trigger a labelling obligation under the AI Act. For image, audio and video content, the AI Act does not separate human-made content from AI-made content. The deciding criterion is whether the result meets the definition of a deepfake.
A deepfake, under Article 3(60) of the AI Act, is image, audio or video content generated or manipulated by AI. This content resembles existing persons, objects, places, entities or events. It may falsely appear to a person to be authentic or truthful.
Two parties are responsible for labelling AI content: the provider and the deployer of the AI system. The provider is the party that develops the AI system and places it on the market. The deployer is the company or organisation that uses the AI system in its activities.
The provider of a generative AI system marks its outputs in a machine-readable format under Art. 50(2). Technical solutions include metadata, watermarking and content provenance technologies. Google Cloud describes the obligations of technology providers in its documentation EU AI Act compliance – Google Cloud.
The deployer, i.e. the company or agency, ensures visible disclosure for deepfake content.
|
Obligation |
Who is primarily responsible |
|
Machine-readable marking of AI output (Art. 50(2)) |
Provider of the AI system |
|
Visible disclosure for deepfake content (Art. 50(4)) |
Deployer of the AI system (company, organisation) |
A watermark or metadata from an image generator does not replace the deployer's obligation to inform the audience about deepfake content.
Contact us via our web form and we will be happy to take care of it.
The deepfake test consists of three conditions that must all be met at once. AI created or modified the content, the content resembles something that exists, and the result may come across as authentic. Content that fails even one condition is not a deepfake.
Content created without AI does not fall under Article 50 of the AI Act. Content created or modified with AI moves on to condition 2.
Besides face swaps, the deepfake definition also covers existing objects, places, entities and events. Condition 2 is met by a visual that resembles:
The third condition is decisive. The criterion is the impression the visual creates in the viewer, not the author's intention. European regulation aims to let audiences tell authentic content from synthetic content. Euractiv describes how the EU's AI transparency rules are changing the content shown online.
Content that meets all three conditions and creates a false impression of authentic reality is a deepfake and must be labelled.
Before publishing an AI visual, a marketing team checks three questions: whether it depicts a real person, whether it depicts an existing place or event, and whether the product matches reality.
A realistic AI person calls for closer scrutiny when viewers could take them for a specific real individual. A stylised illustration, drawing, 3D character or fantasy figure does not create this impression. With these, viewers do not expect a photographic record of reality.
Example: a cartoon AI character presenting a product is assessed differently from a realistic video. In such a video, a well-known personality appears to endorse a product, even though they never made any such recording.
A photorealistic AI photo of an existing shop after a refurbishment that never took place gives customers the impression of an authentic photograph. A made-up interior scene does not depict a specific flat, hotel or shop. A photorealistic AI scene is therefore not automatically a deepfake.
This question goes beyond the AI Act. Marketing communication is also subject to consumer protection rules and the ban on misleading commercial practices.
Recolouring a sofa to a blue variant the manufacturer actually offers is a permissible technical edit. The risk of misleading advertising arises with these AI interventions:
A “created with AI” label does not make a false advertising claim legal.
AI edits that do not change the meaning of the reality depicted do not need a label. Most everyday design and advertising workflows sit outside the deepfake regime. Typical AI edits with no labelling obligation:
Whether a background swap needs a label depends on the context of the final visual. A product placed in a generic, made-up interior does not depict existing reality. A product placed by AI into a specific, existing hotel creates the impression that it is really there, and may show the hallmarks of a deepfake.
A correct AI disclosure informs the user clearly and distinguishably, at the latest at the first exposure to the deepfake content. The wording reflects the actual extent of AI involvement:
The disclosure does not belong in the general terms and conditions or somewhere reachable only after several clicks.
The EU icons for labelling AI content are the European Commission's official set of visual marks. The set contains three types of icon:
Each icon comes in a black, a white and a 50% transparent version. The European Commission describes the use of the icons as voluntary. An icon on its own does not guarantee compliance with Article 50 of the AI Act.
The European Commission recommends placing the icon visibly at the user's very first contact with the content. The icon also stays in place when the content is shared further or downloaded. As an example for the Partially AI-Modified icon, the Commission cites an authentic photo of an empty flat that AI has virtually furnished (EU Icons for labelling AI-generated content).
Evidently artistic, creative, satirical or fictional content is subject to an adapted transparency obligation. The obligation does not disappear. The deployer adapts the form of the disclosure so that it does not hamper the display or enjoyment of the work. A film or artistic work therefore does not need a large warning across the image.
Google Ads and Meta Ads label AI content with their own mechanisms, independently of the AI Act. Platform labelling and the legal obligation under the AI Act are two separate levels. In online advertising, three situations can arise:
With Google Ads, a marketing team distinguishes the legal obligation under the AI Act from the platform's technical and transparency mechanisms. Google is developing its own mechanisms for labelling and tracking content created or modified by generative AI.
No deepfake is created when AI makes these edits:
Google may apply its own AI content labels, metadata and technical signals regardless of the legal deepfake threshold. A marketing team therefore checks two things:
The answers to these two questions may differ.
Meta Ads applies the same principle to ads on Facebook and Instagram. Meta uses its own AI info label and technical mechanisms to identify content that generative AI has created or significantly modified.
Not necessarily. A technical edit of an existing photo differs from an intervention in which AI creates:
Meta may apply its own transparency system even to content that does not reach the deepfake threshold under the AI Act.
Yes, Meta can identify AI creatives made in an external AI generator. Modern systems use metadata, content provenance standards and other technical signals. A company should therefore not assume that Facebook or Instagram will treat an external AI image as an ordinary photo.
|
Situation |
AI Act |
Google Ads |
Meta Ads |
|
Colour correction using AI |
Generally no label |
Possible technical AI signals |
For a minor edit, generally no visible AI label |
|
Removing a small object |
No label if it does not change the meaning of reality |
Depends on how the asset was created |
Depends on the extent of the edit |
|
Generic AI scene |
Not a deepfake if it does not depict existing reality |
The platform may record the AI origin |
The platform may label the AI content |
|
Realistic manipulation of an existing person |
Possible deepfake → label |
A platform label does not replace the legal obligation |
A platform label does not replace the legal obligation |
|
Fake photo of an existing place or event |
Possible deepfake → label |
Separate legal assessment |
Separate legal assessment |
|
AI-generated product feature that does not exist |
Risk of misleading advertising |
Risk of breaching advertising policies |
Risk of breaching advertising policies |
Marketing agencies check AI creatives on three layers: the AI Act, advertising and consumer law, and the ad platform's policies. Passing one layer does not guarantee passing the other two.
Contact us via our web form and we will be happy to take care of it.
A legal or compliance assessment is needed in situations that go beyond what a designer or PPC specialist can check:
In these cases, the AI Act overlaps with the rules on advertising, consumer protection, personality rights, copyright and personal data protection.
The transparency obligations of Article 50 of the AI Act apply from 2 August 2026. For relevant infringements, penalties for businesses reach up to €15 million or 3% of total worldwide annual turnover, whichever is higher (Art. 99(4)).
The European transparency framework is increasing the number of AI labels, watermarks and content provenance notices that users see online. Euractiv analyses the impact of the AI transparency rules on how people use the internet.
|
Deadline or area |
Meaning |
|
2 Aug 2026 |
Art. 50 transparency obligations start to apply |
|
Deepfake |
Visible notice of content created or manipulated by AI |
|
EU AI icons |
Voluntary tool for transparent labelling |
|
Google Ads |
Its own technical and transparency mechanisms |
|
Meta Ads |
Its own AI info label and AI content detection |
|
Penalty |
Up to €15 million or 3% of worldwide annual turnover (whichever is higher) |
A company workflow does not need a complex approval process for every photo opened in an AI tool. A six-point checklist takes care of the checks:
Using AI does not create an obligation to label every visual. A deepfake does require a label. Google Ads and Meta Ads apply their own AI labels, even beyond the legal obligation under the AI Act.
So the key question for a marketing team is not “Did we use AI?”. The key question is: “What did AI change in the final visual, and what impression will it create for the customer?”
This article is for information only and does not replace individual legal advice.
The latest trends, proven practices and analyses from the digital world. Find out how to take your technology and marketing growth further.
And don't forget to check out Invelity Extras – you'll find a marketing glossary, useful tools and plugins to make your work easier!
Invelity Extras