TL;DR
- Since 2 August 2026, the transparency rules in Article 50 of the EU AI Act apply across all 27 member states.
- Deepfakes must be labeled: realistic AI images, video or audio that resemble real people, places or events must be disclosed as AI-generated or manipulated.
- Most marketing text does not: the text rule covers publications that inform the public on matters of public interest, and it falls away when a person reviews the text and takes editorial responsibility.
- Light edits are fine: filters, color correction, noise removal, upscaling, and using AI for ideas, scripts or captions do not turn content into a deepfake.
- Platforms go further: YouTube, Meta and TikTok have their own AI labels and apply them to personal and business accounts alike.
- Fines: breaches of Article 50 can cost up to €15 million or 3% of global annual turnover, with lower caps for small companies.
This explainer summarizes official sources. It is not legal advice.
Contents
- What Article 50 requires
- What counts as a deepfake
- Edits that need a label, and edits that don’t
- AI-written text, ads and books
- How to label: the EU icons
- What YouTube, Meta and TikTok require
- Who is responsible, and the fines
- Bulgaria and the national rules
- FAQ
- Vyara’s take
Generative AI is now part of everyday marketing: product photos, ad visuals, voice-overs, scripts. Since August 2026, the EU has clear rules on when audiences must be told that content was made or changed with AI. The short answer to “do I have to label it?” is: yes, but not every time.
What Article 50 requires
Article 50 of the AI Act splits the duty between two groups:
| Who | Called in the Act | What they must do |
|---|---|---|
| Companies that build AI tools, such as OpenAI, Anthropic, Google or Midjourney | Providers | Mark AI outputs in a machine-readable way, for example with watermarks or metadata, so they can be detected |
| Businesses, agencies, freelancers and creators who use AI tools to publish content | Deployers | Visibly disclose deepfakes and certain AI-generated text to the audience |
The obligations apply from 2 August 2026. The Digital Omnibus on AI, adopted in July 2026, gives providers of generative AI systems already on the market before that date until 2 December 2026 to comply with the machine-readable marking duty. The Commission published guidelines on Article 50 on 20 July 2026.
What counts as a deepfake
Article 3(60) of the AI Act defines a deepfake 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.
Two elements matter:
- Resemblance: real people, their face, body or voice, or real places and events.
- An impression of authenticity: the way the content is presented makes the viewer believe it really happened, or that the person really said or did it.
Typical deepfakes include a real person shown saying something they never said, a cloned voice, a face swap, or a realistic disaster, protest or accident that never took place.
Content that is clearly fantastical or cartoonish, such as an animation or a person riding a unicorn, is not a deepfake. For evidently artistic, creative, satirical or fictional works, the duty is lighter: the existence of AI content must still be disclosed, but in a way that does not hamper the display or enjoyment of the work, for example in the description or the opening or closing credits.
Edits that need a label, and edits that don’t
| Usually needs a label | Usually no label needed |
|---|---|
| Changing what a real person does or says | Beauty filters, color and light correction |
| Synthetic voice or altered speech of a real person | Background blur, noise removal, sharpening, upscaling |
| Face swap or a heavily changed appearance | AI help with ideas, scripts, structure or titles |
| Fake events added to real footage | Resizing images, auto-generated captions |
| A new background or main object created with generative AI | AI-edited ad copy on authentic visuals |
A practical test for product and property photos: if AI only makes a real photo look better, it is a cosmetic edit. If it adds things that do not exist, such as furniture in an empty apartment, a view that is not there or features the product does not have, the image can mislead buyers. The EU’s own icon guidance uses an empty apartment furnished with AI as an example of partially AI-modified content.
Photorealistic AI people in ads need particular care. An AI-generated doctor recommending a product is exactly the kind of content the rules were written for.
AI-written text, ads and books
The text rule in Article 50(4) covers AI-generated text published to inform the public on matters of public interest. It does not apply when the text has gone through human review or editorial control and a person or company holds editorial responsibility for it.
In practice:
- Standard ad copy for products and services is not a public-interest publication. Platforms treat AI copywriting as production help.
- Articles, blog posts and books written with AI do not need a label if the author reads, edits and fact-checks the text and publishes it under their name.
- Fully automated publishing with no human in the loop is where the label becomes mandatory, for example sites that generate and publish articles automatically.
- Political and social-issue ads have stricter platform rules. Meta, for example, asks advertisers to disclose digitally created or altered content in these ads.
How to label: the EU icons
The EU has published official icons for labeling AI content, in black and white versions:
- Basic AI icon: when AI was involved and a custom text label is used.
- Fully AI-generated: content made entirely with AI.
- Partially AI-modified: existing human-made content changed with AI.
The label must be clearly visible at the latest when someone first sees the content, free of overlapping elements, and it should stay visible when the content is reshared or downloaded. The icons themselves are optional; the labeling duty is not. Detailed placement rules for images, video and audio are in the voluntary Code of Practice on Transparency of AI-Generated Content, published on 10 June 2026 and signed by about 190 companies and organizations by the end of July 2026.
What YouTube, Meta and TikTok require
- YouTube: creators must disclose realistic content that is generated or meaningfully altered, such as a real person appearing to say something they didn’t. AI help with scripts, thumbnails, titles or captions does not need disclosure. YouTube says disclosure does not limit a video’s audience or its eligibility to earn money, but creators who consistently fail to disclose risk content removal or suspension from the YouTube Partner Program.
- Meta: ads made or significantly edited with Meta’s generative AI tools get an AI info label. When the tools add a photorealistic AI person, the label appears next to Sponsored. Meta says it will also detect ads made with third-party AI tools through industry-standard signals.
- TikTok: requires creators to label realistic AI-generated content and automatically labels AI content carrying C2PA Content Credentials metadata.
Platform rules apply to personal accounts too. Posting a deepfake publicly is not private or household use, so a non-commercial profile does not remove the duty. Separately, an EU directive from 2024 requires member states to criminalize sharing intimate or sexual deepfakes of real people without their consent.
Who is responsible, and the fines
- The business being advertised is the deployer and carries the main legal responsibility to disclose.
- Agencies and freelancers act on the business’s behalf. The business can seek compensation from them under their contract, and platforms can block an agency’s own ad account.
- Platforms must give advertisers tools to label content, and can label, remove or block content and accounts that break their rules.
- AI tool providers are responsible for machine-readable marking, not for how an advertiser later uses or strips the content.
Under Article 99, breaches of Article 50 can be fined up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups, the lower of the two applies. Prohibited AI practices carry higher fines, up to €35 million or 7%.
The rules also apply to companies outside the EU when their AI content is offered or aimed at people in the EU, and to EU-based businesses advertising abroad.
Bulgaria and the national rules
Each member state appoints its own market surveillance authorities, while the EU AI Office oversees general-purpose AI models. In Bulgaria, a bill on the use and development of artificial intelligence was submitted to the National Assembly on 24 September 2026 by a group of MPs and assigned to committees. According to reporting on the draft, oversight would be split by sector between regulators such as the Consumer Protection Commission and the Council for Electronic Media. As of mid-October 2026, it is a proposal, not an enacted law. The EU rules already apply regardless.
FAQ
Do I have to label an ad visual I improved with ChatGPT?
Not if AI only improved a real photo, for example light, color or sharpness. If it added elements that do not exist in reality, the image can mislead and should be labeled.
Do I have to label ad copy written with AI?
Generally no. Standard commercial ad copy is treated as production help. Political and social-issue ads have stricter platform rules.
Do I have to label a blog post or a book written with AI?
Not if you review, edit and fact-check the text and publish it under your own name. Fully automated publishing without human review does need a label.
Does the rule apply to personal accounts?
Yes. Publishing a realistic deepfake online exposes it to the public, and YouTube, Meta and TikTok apply their AI labels to personal and business accounts alike.
What about parody and satire?
The duty is lighter but does not disappear. A short note in the description, caption or credits is enough, as long as it does not spoil the work.
Will an AI label hurt my reach?
YouTube says disclosure does not limit a video’s audience or monetization. Audience reaction is a separate question: in research released by The Harris Poll at Cannes Lions 2026, 78% of consumers said AI makes ads feel less authentic.
Vyara’s take
“Labeling is not the problem. Misleading people is. Use AI to make real things look better, not to show things that don’t exist. If you would feel uncomfortable when a customer found out how an image was made, label it. And keep a human in the loop: it protects you legally and it makes the work better.”
Further reading: Do you have to label AI content? – Vyara Stefcheva’s personal take on stefcheva.com.
Sources: AI Act, Regulation 2024/1689 · Article 50 on the AI Act Service Desk · Article 99 on the AI Act Service Desk · Digital Omnibus on AI, Regulation 2026/1744 · Cuatrecasas on the Digital Omnibus · Commission guidelines on Article 50 · Code of Practice on AI-generated content · EU icons for AI content · Directive 2024/1385 · Bulgarian AI bill · YouTube disclosure policy · Meta on AI transparency in ads · NBC News on TikTok labels · The Harris Poll at Cannes Lions 2026
About the Author
👉Want guest posting?🪙Subscribe to the @egorithms Telegram