The rules for labeling AI content ceased to be a recommendation on August 2, 2026, when Article 50 of the AI Act took effect.
The regulation operates on two levels: a machine-readable marker embedded in the content itself, and a visible label for people. The first is primarily the responsibility of generative system providers, while the second falls on anyone who publishes certain types of AI-generated content. Confusion over these roles is the most common reason why websites end up doing more or less than is necessary.
Съдържание
The Key Points of the AI Act in Six Points
The summary is compiled into a short list.
- Regulation (EU) 2024/1689 classifies systems by risk: prohibited practices, high-risk systems, systems subject to transparency requirements, and all others without specific obligations
- The bans take effect in February 2025, the rules for GPAI models in August 2025, and the transparency requirements under Article 50 and actual supervision on August 2, 2026.
- High-risk systems listed in Annex III will become effective on December 2, 2027, and those embedded in regulated products will become effective on August 2, 2028.
- Responsibilities depend on the role: the system supplier applies machine-readable labels, while the integrator applies visible labels
- Fines can reach up to 35 million euros or 7% of global annual revenue for prohibited practices
- The Transparency Code of June 10, 2026, is voluntary but implements the legal obligations and is recognized by the Commission as evidence of compliance
The full text of the regulation is available through the Commission’s AI Act Explorer, including in Bulgarian. However, Article 50 is the one that is primarily relevant to the day-to-day operations of a website.
What types of content require a visible label?
A visible label is required in only two cases. The first is a deepfake: an image, audio, or video that resembles real people, objects, places, or events and may mistakenly appear authentic. The second is AI-generated text that informs the public on matters of public interest, unless it has been edited by a human who assumes editorial responsibility. Everything else, including standard AI illustrations and product descriptions, does not require a visible label.
The exceptions for creative content are less restrictive, but they do not cover advertising. In its guidelines under Article 50, the Commission explicitly states that AI videos featuring realistic simulated people recommending a product constitute deepfake advertising and are subject to labeling.
How to Apply the Visible Label
The label must be clearly visible upon first encountering the content. For text, this means at the beginning, next to the author’s name, or before the title; a label in the footer does not meet this requirement. For images and videos, the label must be embedded directly in the content or displayed via an equivalent element in the interface, without other elements obscuring it. The wording must be in plain language and accessible to screen readers.
The EU has also published three free icons for AI-generated content: one for fully generated content, one for partially modified content, and a general default icon. The files are available in SVG and PNG formats, in black, white, and semi-transparent versions, with no attribution required. The icon alone is not sufficient and works best when accompanied by a short text label.
EU Icons
The icons come in four versions: black, white, black with 50% transparency, and white with 50% transparency. You can download ZIP files containing all the icons in all versions in SVG and PNG formats.
The icons underwent user testing, and the results influenced their design. It is noteworthy that performance improved across all metrics when the main icon was accompanied by a text label (e.g., “modified”).
| Icon | When is it used? | Examples |
Main Icon | When artificial intelligence has been used to create fake content (images, audio, video) or published text, or when a personalized text label or interactive overlay has been added. | A deepfake video with the text label “voices generated with,” followed by the main icon |
Generated entirely by artificial intelligence | When all fake content (images, audio, video) or text is generated entirely by artificial intelligence without any human input or human editorial oversight (except for prompts) | Deepfake videos featuring politicians or fictional events, generated entirely by artificial intelligence; Music or art entirely composed by artificial intelligence;* News summaries generated by artificial intelligence. |
Partially modified using artificial intelligence | When it existed previously, the human-generated content was partially modified using artificial intelligence, turning it into deepfake text or text on topics of public interest. | A person’s face in an authentic photo was replaced with that of a politician using artificial intelligence. Authentic photos of an empty apartment were generated using artificial intelligence. |
*May be used with limited disclosure requirements for artistic, creative, or satirical works
The use of these EU icons is optional, but the labeling requirements under Article 50 of the Artificial Intelligence Act are not. The use of these icons alone does not establish compliance with the law. Implementers remain responsible for ensuring that any disclosure complies with the requirements of Article 50 of the Artificial Intelligence Act. Signatories to the Code of Practice on the Marking and Labeling of AI-Generated Content must duly implement the measures set forth therein.
The Machine-Readable Marker in Practice
The Code of Good Practice requires providers to include at least two parallel layers: digitally signed metadata with a C2PA-compliant timestamp and an invisible watermark within the content itself. Google implements the second layer using the SynthID watermark, while OpenAI and Adobe sign C2PA metadata. For short or heavily transformed texts, the code allows for fingerprinting or logos as a fallback approach. The two layers complement each other: the metadata is easy to read, while the watermark is more resistant to transformations.
The weak point is data loss along the chain. Compression and conversion to WebP often erase embedded metadata, so a marker that survived in the original may be missing from the published version.
Websites and Online Stores: A Practical Guide
For a typical WordPress site or WooCommerce store, responsibilities are divided by component.
- Chatbot: A message such as “You are chatting with an AI assistant” when the window opens; most SaaS chatbots now offer this as a setting, but check to see if it’s enabled
- Product images: The generated illustrations and banners are not deepfakes and do not require a label; however, AI videos featuring realistic “satisfied customers” are exactly that
- Product Descriptions: AI-generated product text is not a matter of public interest and does not require labeling, but misleading claims remain a violation of consumer rights
- Blog and News: AI-generated articles on topics of public interest that have not been edited by a human are labeled as such at the beginning; if a human has edited the article, document the review to qualify for the exception
- Media Library: Keep the original files with their metadata and make sure that image optimization doesn’t overwrite them
- Legal Page: Describe the use of AI in the site’s terms and conditions
A visible label on an AI image can be added with just a few lines of code. All it takes is a custom field in the attachment and a single filter.
The regulation will not remain in its current form: the guidelines under Article 50 were updated on July 20, 2026, and the enforcement of sanctions is just beginning.
Learn more about the fines under the AI Act.
Frequently Asked Questions
-
When does an AI-generated image in an online store require a visible label?
Only when it is a deepfake: it realistically depicts people, places, or events and could be mistaken for the real thing. Ordinary AI illustrations, banners, and product visuals do not require a label.
-
Where should the label be placed in AI-generated text?
At the beginning of the publication, next to the author’s name or before the title, so that it is visible upon first viewing the content. Placing it only in the footer does not meet the requirement.
-
Are EU icons for AI content mandatory?
No, the icons are a free option with no attribution requirement. Only the designation itself, as required by Article 50, is mandatory; the icon serves as a supplement to a brief text label written in plain language.
-
Who is responsible for the machine-readable marker in AI content?
The provider of the generative system, which embeds metadata and a watermark. The implementer must not remove the marker during processing and must visibly label deepfakes and certain AI-generated texts.
