Your chatbot signs off as “Ania from customer service”? From 2 August 2026, that could cost up to EUR 15 million. But the duty to label AI content does not apply to everything you generate. It applies in exactly four situations.
Article 50 of Regulation (EU) 2024/1689 (the AI Act) imposes specific transparency obligations on providers and deployers of AI systems. Breaching these obligations may result in an administrative fine of up to EUR 15,000,000 or 3% of total worldwide annual turnover, whichever is higher. For SMEs and startups, the lower of the two amounts applies.
This article explains:
- Four situations in which you must label AI content
- A simple test to help determine whether your content needs a label
- Exceptions: what you do NOT need to label
- Practical implementation steps you can complete before 2 August 2026
Key concepts you need to know
Before we get into the details, determine your company’s role in the AI Act ecosystem. The scope of your obligations depends on it.
| Term | Meaning | Example |
|---|---|---|
| AI system provider | An entity that develops an AI system, or has one developed, and places it on the market under its own name | A company that built a chatbot and sells it as a product |
| Deployer | A company or institution that uses an existing AI system under its authority, not for personal purposes | An e-commerce business using Midjourney for product graphics |
| Deepfake | An image, audio, or video generated or manipulated by AI that a viewer could mistakenly consider authentic | A virtual model in an advertising campaign, or a synthetic voice in a commercial |
| Synthetic content | An image, audio, video, or text generated by an AI system | A product render, or an article written by a language model |
| Editorial control | Substantive human review of content, where a specific person or company takes responsibility for publication | An editor reviews and endorses an article written with AI assistance |
One point to remember: it is easy to move from the role of deployer to that of provider. It is enough to put your own logo on a system, modify how it works, or change its intended purpose. Each of these changes means a broader set of obligations.
Four situations in which you must label AI content
Article 50 of the AI Act defines four separate categories of transparency obligations. Each concerns a different type of content and a different responsible party.
1. Interaction with an AI system: chatbots and voicebots
Who is covered: providers of AI systems intended for direct interaction with natural persons.
What the obligation requires: the system must inform the user that they are interacting with AI no later than the first interaction. The notice must appear at the start of the conversation, not afterwards.
What this means for your company:
- A website chatbot signed “Ania from customer service” breaches the rule. It must clearly say that it is an automated assistant.
- A call-center voicebot must say at the start of the call that the caller is speaking with an AI system.
- Remove elements that suggest the user is speaking with a human: human names, operator profile photos, and avatars that look like real people.
Exception: disclosure is not required where the fact that a person is interacting with AI is obvious to a “reasonably informed, observant, and circumspect natural person,” taking the context into account. But this exception is narrow. If you are unsure whether your chatbot is “obvious,” it is safer to label it.
2. Technical marking of synthetic content: metadata and watermarks
Who is covered: providers of AI systems that generate synthetic images, audio, video, or text.
What the obligation requires: system outputs must be marked in a machine-readable format so they can be detected as artificially generated or manipulated. This may involve metadata, watermarks, cryptographic methods, or digital fingerprints.
Who is responsible: this is the tool provider’s obligation, for example, the creator of an image generator, not yours as the user. But if you modify the system enough to become a provider, the obligation shifts to you.
Transition period: the Digital Omnibus on AI (Regulation 2026/1744, which entered into force on 27 July 2026) gave providers of generative systems placed on the market before 2 August 2026 until 2 December 2026 to implement technical watermarking. Systems placed on the market from 2 August 2026 must apply a mark from day one.
Exception: systems used to assist with “standard editing,” meaning typical editorial actions that do not create new synthetic content, such as cropping, color correction, or sharpening.
3. Deepfakes: realistic images, voices, and video
Who is covered: deployers, meaning your company if it publishes content generated or manipulated by AI.
What the obligation requires: you must disclose that the content was artificially generated or manipulated. The label must be:
- Visible on first contact with the content, not hidden in terms and conditions
- Placed next to the specific content; a general statement on the website is NOT enough
- Persistent: it must remain when content is shared, downloaded, displayed as a thumbnail, or posted on social platforms
Simple test: could someone mistake this content for a real photo, recording, or video? If yes, label it.
The form of the label depends on the medium:
| Content type | How to label it |
|---|---|
| Graphic or image | A legible caption, for example, “AI-generated image,” or the official EU icon with the acronym “AI” |
| Video | A fixed overlay visible throughout the entire video |
| Audio or voice | A spoken notice, since audio alone has no image on which to place a label |
A platform label on Instagram, TikTok, or LinkedIn does NOT replace your duty. The platform is not the deployer of your content. If the platform’s “AI” label disappears when the file is downloaded, you must embed the label in the content itself.
Artistic exception: for a deepfake in a clearly artistic, satirical, or fictional work, the disclosure may be less prominent, for example in the end credits, but the duty is not waived. An advertisement pretending to be art does NOT qualify. A virtual model in an ad campaign or AI-generated customers praising a product are commercial content and require full disclosure.
4. AI-generated text on matters of public interest
Who is covered: deployers, meaning entities that publish AI-generated or manipulated text to inform the public about matters of public interest.
“Matters of public interest” include health, safety, the environment, politics, the economy, investor communications, and ESG reports.
Exception: the text has been reviewed by a human or subjected to editorial control AND the publishing entity has editorial responsibility. Both conditions must be met together.
This means that if your editor substantively reviews an AI-assisted article about cybersecurity and endorses it, there is no duty to label it. But merely “reading” the text is not enough. Editorial control means substantive verification and acceptance of responsibility for the content.
Ordinary content marketing, such as product guides, feature descriptions, or blog posts about your product, usually does not concern matters of public interest and is outside the scope of this duty.
Contact us if you need help classifying your content.
What you do NOT need to label: exclusions
Not every AI-generated item needs a label. Create an internal list of exclusions so your team does not label everything blindly.
| Content type | Label required? | Why? |
|---|---|---|
| Abstract AI graphics (icons, charts, infographics) | No | Viewers will not mistake them for real photographs |
| Cropping, color correction, or sharpening a photo | No | Standard editing does not create new synthetic content |
| A product guide written with AI | No | Usually does not concern matters of public interest |
| Content generated AND published before 2 August 2026 | No | No retroactive labeling duty |
| A realistic 3D product render made with AI | Yes | Viewers may mistake it for a real photograph |
| Virtual home staging in a property listing | Yes | An AI-generated realistic interior may be mistaken for a real one |
| A synthetic voice in an advertising spot | Yes | Deepfake audio |
| A chatbot that does not say it uses AI | Yes | No notice that the user is interacting with an AI system |
Rule for older content: materials generated AND shared before 2 August 2026 are exempt. But a render created in July 2026 and published in September 2026 must be labeled because its first publication is after the cutoff date.
Who is responsible when an agency or freelancer creates the content?
This is one of the most common questions. A creative agency produces a campaign using AI content. Who gets fined if the material is not labeled?
The answer depends on who is the deployer: who actually decides to use the AI system and publishes the result under their authority.
| Scenario | Who is the deployer? | Who is responsible for labeling? |
|---|---|---|
| The agency independently chooses an AI tool and creates content | The agency | The agency |
| The client instructs the agency to use a specific AI tool and controls the result | The client | The client |
| No contractual arrangements and unclear division of roles | Risk of a dispute | Both parties risk a fine |
There is no case law on this issue. State directly in the contract who is the deployer, who is responsible for labeling, and who ensures the label survives cropping and editing.
Five implementation steps before 2 August 2026
The duty can be implemented in one sprint. Here is a practical sequence.
Step 1. Inventory AI content in the company. Identify every place where AI content is created: marketing, e-commerce, customer service, and HR. Determine the content type: image, video, audio, text, or chatbot. For each item, apply a simple test: could the audience mistake it for something authentic?
Step 2. Exclude content outside the duty. Create a clear list of exclusions: obvious illustrations, icons, charts, abstract graphics, ordinary company text, and content published before 2 August 2026. This prevents the team from wasting time labeling material that does not require it.
Step 3. Add a labeling mechanism in your CMS. Create a component or “AI” badge, using the official EU icon or a legible caption, and display it next to every item that needs a label. For downloadable and shareable files, the label must be embedded in the file itself, as an overlay on an image or video, not only in the website interface. Test what happens when materials are downloaded from Instagram, TikTok, and LinkedIn. If the label disappears, embed it permanently.
Step 4. Configure chatbots and voicebots. Add an opening notice: “You are speaking with an automated assistant.” Remove elements that suggest a human is responding.
Step 5. Update agreements with agencies and freelancers. State directly who is the deployer, who is responsible for labeling, and who ensures the label remains after the material is edited.
Penalties: what is the real risk?
| Type of violation | Maximum fine | Rule for SMEs and startups |
|---|---|---|
| Prohibited AI practices (Article 5) | EUR 35 million or 7% of turnover, whichever is higher | The lower of the two amounts |
| Transparency obligations (Article 50) | EUR 15 million or 3% of turnover, whichever is higher | The lower of the two amounts |
| Providing false information to authorities | EUR 7.5 million or 1% of turnover, whichever is higher | The lower of the two amounts |
Fines for failure to label AI content are the second-highest category of AI Act sanctions, comparable in scale to GDPR fines (up to EUR 20 million or 4% of turnover).
SME and startup mechanism: Article 99(6) of the AI Act reverses the rule. Instead of the higher amount, the lower one applies. For a startup with turnover of EUR 500,000, the fine will not exceed EUR 15,000 (3% of turnover), not EUR 15 million. This is a hard ceiling that the authority cannot exceed.
When setting a fine, the authority takes into account, among other things:
- The nature, seriousness, and duration of the violation
- The number of people harmed and the scale of the damage
- The technical and organizational measures applied
- The level of cooperation with the authorities
- Previous violations
In Poland, the supervisory authority will be the Commission for the Development and Safety of Artificial Intelligence (KRiBSI), supported by the Ministry of Digital Affairs. The draft AI Systems Act provides for a settlement allowing an exceptional reduction of sanctions. The fine may be reduced by as much as 90% if the entity implements the Commission’s recommendations following a warning and no harm to health has occurred.
At present, there are no precedents for fines for violating Article 50. Authorities are focusing on explaining the obligations, but this will change after 2 August 2026.
AI content labeling and other regulations
Meeting the AI Act’s transparency obligations does not automatically mean compliance with other laws. Recital 137 of the AI Act states this expressly.
Be aware that obligations may overlap:
- GDPR: if AI content uses a specific person’s image (personal data, potentially biometric data), you need a separate legal basis for processing. Labeling the content as AI-generated does not replace consent to process personal data.
- Copyright and image rights: a synthetic voice or image of a person is a personal interest protected under the Civil Code. The Supreme Court judgment of 7 March 2023 (II CSKP 659/22) confirms that a voice is protected as a personal interest.
- Unfair Competition Act: failing to label realistic AI content in an advertisement may constitute misleading advertising.
Technical standards: what to choose
The European Commission has not named one mandatory technical standard for machine-readable labeling. It promotes layered marking, and two standards operate as de facto solutions:
| Standard | Description | Use |
|---|---|---|
| C2PA (Content Credentials) | An open cryptographic standard for verifying content provenance and editing history | Images, video, and documents; supported by, among others, Adobe and Microsoft |
| IPTC | A metadata standard for media content | Photos and graphics; widely used in media and photo agencies |
The Code of Practice on Transparency of AI-Generated Content is voluntary, but following it may help demonstrate compliance with Article 50.
How we can help your company
The duty to label AI content takes effect on 2 August 2026. Implementation requires not only technical CMS configuration but, above all, legal classification of content: distinguishing a deepfake from an obvious illustration, determining provider and deployer roles, and reviewing agency agreements.
We support technology companies, e-commerce businesses, and startups with:
- Classifying AI content: assessing which materials are subject to labeling and which fall within exceptions
- Determining roles: identifying who in your content-production chain is the provider and who is the deployer
- Reviewing agreements with creative agencies, freelancers, and AI tool providers, including responsibility clauses for labeling
- Analyzing overlapping duties under GDPR, copyright, and image rights
- Editorial-control procedures: documenting the review process for AI-generated texts that removes the labeling duty
Summary: what to do before 2 August 2026
The duty to label AI content applies in four specific situations, not to everything you generate. Most of your AI content probably does not need a label. But when a label is required, it must be done properly because the fine may reach EUR 15 million or 3% of turnover.
Three things to do now:
- Inventory your company’s AI content and apply the simple test: could the audience mistake it for something real?
- Implement a labeling mechanism in your CMS and test whether labels persist on social platforms
- Update agreements with agencies and freelancers to establish who is responsible for labeling
If you need help classifying content, reviewing agreements, or implementing labeling procedures, contact us.
Frequently asked questions
We use Midjourney for graphics on our company blog. Do we have to label every image as AI-generated?
No. Apply the simple test: could a viewer mistake the image for a real photograph? If it is an abstract illustration, icon, or infographic, you do not need to label it. If it is a realistic image that looks like a photograph, label it. Cropping, color correction, and sharpening do not create a labeling obligation.
A creative agency is producing an AI campaign for us. Who faces the consequences if the content is not labeled?
It depends on who is the deployer, meaning who actually decides to use the AI system and publishes the result. If the agency independently chooses the tool and creates the content, the agency is responsible. If you instruct the agency to use a specific tool and control the result, you are responsible. There is no case law on this issue, so state the division of roles directly in the agreement.
We write articles with ChatGPT but always edit them. Is that “editorial control” under the AI Act?
Editorial control requires both of these conditions: (1) substantive human review of the content and (2) a specific person or company taking editorial responsibility for publication. Merely reading the text is not enough. If an editor substantively reviews and edits the content and endorses it under their name or the company’s name, the conditions for the exception are met. But this applies only to text about matters of public interest. Ordinary content marketing and product guides are usually outside this duty.
Is a footer notice saying “Some content on this website was created with AI” enough?
No. The label must appear next to the specific content and be visible on first contact with it. One general notice for the entire website does not meet Article 50 requirements.
We have hundreds of old AI images on our website. Do we need to review the whole archive?
Content generated AND shared before 2 August 2026 is exempt from the duty. You do not need to label it retroactively. But if an image was generated in July 2026 and published in September 2026, it must be labeled because its first publication was after the cutoff date. Review materials scheduled for publication from August 2026 onward.
We make 3D product renders with AI. Are they deepfakes or obvious illustrations?
It depends on how realistic they look. If a render looks like a professional product photograph and viewers could mistake it for one, treat it as a deepfake and label it. If it is clearly stylized and nobody would mistake it for a photograph, the duty is unlikely to arise. If in doubt, labeling is safer. Its cost is minimal compared with the risk of a fine.
