The essentials in 60 seconds / tl;dr
- Not every AI-generated item used commercially automatically needs a public AI disclosure.
- A clear and perceivable disclosure becomes particularly relevant when realistic AI content could be understood as a real person, place, business, object or actual event.
- The Article 50 rule applies from 2 August 2026; the AI literacy obligation has applied since 2 February 2025.
- Technical machine-readable marking is the system provider’s responsibility. The publishing business is responsible for any required clear disclosure.
- Realistic AI voices and video may also need a clear, perceivable disclosure no later than first exposure. For audio, that disclosure can be audible.
- Businesses should still apply human review, clear rights and advertising claims, and establish internal rules for AI use.
Short on time?
Sources: European Commission: Article 50 Q&A · European Commission: Article 50 Guidelines · European Commission: AI literacy Q&A
01
The core rule: using AI alone does not trigger a disclosure duty
Article 50 distinguishes between technical duties for AI system providers and disclosure duties for professional deployers. For businesses using finished tools to create marketing content, the rule is particularly relevant. It applies from 2 August 2026.
Under the AI Act, a is AI-generated or manipulated image, audio or video content resembling existing or plausibly existing persons, objects, places, entities or events and falsely appearing authentic or truthful. It is not limited to fake celebrity videos.
The assessment considers factors such as resemblance, the substantive message, the deployment context and the audience. If the audience clearly does not expect documentary truth, that may weigh against a classification. The assessment remains context-specific.
does not only mean impersonating a real person
AI content
The image, audio or video was generated or substantially manipulated with AI.
Plausible reality
It resembles a person, object, place, business or event that exists, could plausibly exist or could plausibly have existed.
False appearance of authenticity
A person could mistakenly understand the essential content as authentic or truthful.
A completely invented hotel, restaurant, product, business or event may therefore still be relevant. An identifiable real person is not required. The three criteria must be considered together.
02
Pixels, metadata and provenance: what exactly must be present?
Article 50 creates two separate duties. The provider of the generative AI system must enable technical detection. The business publishing relevant content may also have to provide a clear and perceivable disclosure to people. One layer does not replace the other.
What the law itself requires
The AI Act does not prescribe one technology or file format for providers. The result must be machine-readable and detectable, and the solution must be effective, interoperable, robust and reliable as far as technically feasible. A publishing deployer must disclose s clearly and distinguishably no later than first exposure.
The three technical building blocks
Digitally signed metadata
Records in machine-readable form that a file was generated or modified with AI, for example using the open C2PA standard (Content Credentials) backed by Adobe, Google, Microsoft and OpenAI, among others. It can be lost during export or upload, though.
Imperceptible watermark in the image data
Embedded imperceptibly in the pixel structure and intended to survive processing such as compression or cropping. It is not a visible label.
Provenance data and an audit trail
Additional details about the system, provider and time improve provenance.
This two-layer approach comes from the voluntary EU Transparency Code. Non-signatories may use equivalent technical solutions but must be able to demonstrate their effectiveness.
When must the notice be visible or audible?
If an image, audio or video constitutes a , invisible technical marking is not enough. This may also concern a realistic-sounding AI voice or an invented but plausibly real-looking business, product, hotel or event. The disclosure must be understandable and perceivable without specialist tools or an additional click. For audio, that can be an audible notice; for video, it will usually be a clear overlay. The AI Act does not literally require the notice to be permanently burned into the file. For exported files, however, that is usually the most robust option because it survives sharing.
- Clear and perceivable no later than first exposure
- Clear wording such as “AI-generated voice”, “AI-manipulated video” or “AI-generated depiction”
- For video, sufficient contrast, readable size and no obstruction by platform overlays
- For audio, an understandable audible notice, preferably also shown as text in the player
- Directly embedded in the content or presented as an equivalent interface overlay on the content
- Still present after download, export and resharing
- Additional alt text or an ARIA label for assistive technologies
Which EU icons exist and why ours looks different
The European Commission provides three voluntary variants: the basic “AI” icon for content whose creation involved AI, “Fully AI-Generated” for entirely generated content and “Partially AI-Modified” for pre-existing human-made material that was substantially altered with AI. The variants may be used for image, audio, video and text content that is subject to disclosure. There is no separate icon intended exclusively for text. Whether a text must be disclosed at all still depends on the conditions of Article 50.
The images in this article deliberately use a custom, permanently embedded “AI GENERATED” label on a high-contrast background. The labelling icons published by the European Commission may be used voluntarily, but they are not mandatory. Our label is therefore neither an official EU symbol nor a seal of approval or, by itself, proof of legal compliance. It is our attempt to make the disclosure clear, easy to read and visible even after download or resharing.
Sources: European Commission: labelling icons
Who is responsible for what?
System provider
Must ensure the technical machine-readable marking and its detectability; a limited transition applies to older systems. If it relies on third-party marking technology, its own responsibility remains.
Publishing business
Anyone using an AI system professionally under their own authority is a “deployer” under the AI Act. They must assess whether their specific image, audio or video is a and, if so, ensure a clear disclosure. The tool’s metadata or watermark is not enough.
Agency or freelancer
Where a person works under the client’s responsibility and control, the commissioning company will generally remain the deployer. Where the person operates independently, they may be the deployer. The approval responsibility should be clear contractually and operationally.
Platform or export workflow
May technically preserve or remove labels and metadata. That does not automatically relieve providers or deployers. Businesses should test what remains after upload, compression, download and reposting.
A restaurant, garage, agency or freelancer can therefore be a deployer without developing an AI model.
Practical rule for small businesses: archive the original file with existing metadata, do not intentionally remove metadata, integrate the clear notice into the final export and require approval by a named person before publication.
03
AI voices and video: when exactly must they be labelled?
1. Live AI voice or phone bot
When an AI speaks directly with a person, the person must be clearly and distinguishably informed no later than the start of the first interaction that an AI is responding. This covers voice assistants, phone bots and speaking avatars. The duty initially lies with the system provider. A business deploying the service should still check that the notice actually reaches people in its own channel.
2. AI-generated audio file or synthetic voice
Providers of generative AI systems must mark synthetic audio in a machine-readable format and enable it to be detected as artificial. This is a technical mark, not a substitute for any public disclosure that may be required. A generic synthetic narrator is not automatically a deepfake. Realism, context and the audience’s expected perception matter.
3. Realistic voice clone or fake recording
If a real person’s voice is convincingly imitated or a recording sounds like evidence of a real event, a deepfake classification becomes more plausible. The publishing business must then disclose no later than first exposure that the voice was artificially generated or manipulated. For audio, this can be an audible notice before the contribution and should also appear as text in the player.
What about video?
The same basic rule applies to realistic AI video and deepfake video. A clear, easy-to-read disclosure must be perceivable no later than first exposure. A persistent overlay in the video or an equivalent overlay in the player is usually the most robust implementation for export, download and resharing. Metadata alone is not enough for public disclosure.
Which wording works?
- “AI-generated voice”
- “Audio generated with AI”
- “AI-manipulated video”
- “AI-generated depiction”
The AI Act does not prescribe one exact phrase or icon. The official EU icons are optional. What matters is that the disclosure is clear, distinguishable, understandable and perceivable no later than first exposure. The limited transition until 2 December 2026 concerns only the technical marking duty for certain systems already placed on the market before 2 August 2026, not public disclosure of a deepfake.
04
Two fictional advertising images and the difference that matters
Both images were generated specifically for this guide; every business, person, place and event shown is fictional. The AI notice is burned directly into the image files, so it stays in place when the images are downloaded or reshared; the files also declare their AI origin in machine-readable metadata (IPTC digital source type). The examples illustrate typical considerations, not an automatic legal classification. The same content may be assessed differently depending on its caption, platform and audience.
Risk example
A race that looks like authentic local reporting
The image looks like a photograph of a real local race outside a real café. If a business used it to advertise an event or participation that never occurred, viewers could mistake it for authentic evidence. A classification becomes more plausible, and a clear disclosure should be visible no later than first exposure.
Positive example
Obvious fiction, communicated transparently
Foxes selling pastries on the moon would not reasonably be understood as documentary evidence of a real event. A classification is therefore much less likely. The voluntary label still adds clarity and trust without getting in the way of the creative.
Why influencers should look particularly closely
A story may disappear after 24 hours, but that does not remove the relevant publication duties. A personal-looking account may still be used professionally. Sponsored content, affiliate posts, invitations and collaborations should therefore not be treated like private communication where AI content could suggest real experiences, places or events.
05
AI text: advertising is not automatically subject to disclosure
The special disclosure rule for AI text concerns publications intended to inform the public on matters of public interest. The Commission lists areas such as politics, public security, health, environmental protection, consumer safety and relevant economic, scientific or cultural developments.
Ordinary product copy, social captions or menu descriptions therefore do not fall under this specific text rule merely because they are advertising. Closer assessment is needed when advertising communicates claims about consumer safety, health or another matter of public interest.
According to the Commission, text on matters of public interest does not require the AI label where it has undergone genuine human review or editorial control and a person or entity holds editorial responsibility. A spelling or grammar check alone is not enough.
Always verify, regardless of labelling
- Prices, dates and availability
- Addresses, names and contact details
- Product features and performance claims
- Health, safety and sustainability claims
- Sources, figures and quotations
Sources (1)
06
Chatbots and AI assistants: people must be able to tell
Article 50 also covers AI systems that interact directly with people. Anyone writing to a chatbot or voice assistant must be able to tell that an AI is responding. Ensuring this is primarily the system provider’s job. Businesses deploying a chatbot in support or on their website should still check that the notice is actually visible in their specific setup.
An exception applies only where the AI nature is already obvious from the circumstances to a reasonably observant person. When in doubt, a clear notice at the start of the conversation is the simplest solution.
In practice, for website and support chatbots
- Show a notice no later than the start of the first interaction, such as “You are chatting with an AI assistant”
- The bot’s name and presentation do not impersonate a human contact
- When choosing chatbot software, check that this notice is supported and clearly visible
Article 50 also sets information duties for emotion recognition and biometric categorisation. Such systems play no role in most marketing operations and are therefore out of scope here.
07
AI literacy: small marketing teams are included too
The AI literacy obligation has applied since 2 February 2025. Providers and professional deployers should take measures to support the AI literacy of staff and other people operating AI systems on their behalf. Measures should reflect their experience, the use context and the risks involved.
The Commission expressly addresses employees using ChatGPT for advertising copy or translation: the company should inform them about specific risks such as hallucinations. No particular certificate or governance structure is mandated; internal records of training and guidance can be useful.
A lean baseline for small businesses
- List of approved AI tools
- Rules for confidential and personal data
- Checks for facts, rights and visual claims
- Named approval responsibility before publication
- Documented short briefing for staff and contractors
08
The AI Act does not replace other legal checks
Content may not require a public Article 50 disclosure and still be problematic for other reasons.
Misleading advertising
Could the image or copy create a false impression about a product, price, service, event or availability?
People and data
Are real or identifiable people, voices, customer data or other personal information involved?
Copyright
May the source photos, music, copy, designs and other protected material actually be used and modified?
Trade marks and logos
Is there permission for visible brands, sponsor logos, vehicle graphics or business identifiers?
09
One EU regulation, different national practice
The AI Act creates the common EU framework, but practical enforcement of Article 50 and penalties sit mainly with national authorities. National rules on personality rights, unfair competition and media law also continue to matter. An EU-wide marketing process should therefore use one common baseline and add country-specific review for sensitive campaigns. This guide deliberately does not offer a final interpretation of individual national laws.
Cross-border: passport and company location are not the only factors
- A person based or operating in the EU is not outside the scope merely because they are a US or UK citizen.
- Providers and deployers in a third country may also be covered where the output produced by the AI system is used in the EU.
- An EU client or an account aimed at an EU audience is an important signal. The concrete use of the output in the Union matters, not only the billing address.
10
Overview and checklist before publishing
At a glance: what needs labelling when?
| Content type | Labelling required? | What matters |
|---|---|---|
| Realistic AI image, audio or video | Yes, if deceptively real | Clear and perceivable notice where the content could falsely be understood as a real person, place, product or event. For audio, the notice can be audible. |
| Live AI voice or phone bot | Yes, from first interaction | Users must be able to tell no later than the start of the first interaction that an AI is responding, unless this is obvious from the context. |
| Obvious illustration or fantasy visual | No | A voluntary label still adds clarity and trust. |
| Light AI photo editing such as retouching or backgrounds | Usually no | As long as the message and core content of the image are not substantially changed. |
| AI-written advertising and product copy | Usually no | Required only for matters of public interest without editorial responsibility. Always human-check the claims. |
| Website or support chatbot | Yes | The AI must be recognisable no later than the start of the conversation, unless it is obvious from the context. |
The 8-point checklist
Establish provenance
Which tool, source material and substantial AI modifications were used?
Assess realism
Does a person, place, business, object or event look like authentic documentation?
Consider context
What claim is created by the caption, hashtags, campaign and publication timing?
Place the notice
Where disclosure is required or useful: make it clear, perceivable, present by first exposure and retained in the final export.
Preserve technical markings
Do not intentionally remove existing metadata or imperceptible AI markings, and test the platform export.
Verify claims
Human-check prices, facts, names, dates and performance promises.
Check rights
Review people, trade marks, logos, source material and licences.
Record approval
Keep a brief record of the responsible person, date and checks completed.
Simple disclosure wording
AI-generated depictionExample image created with AIAI-generated illustrationOriginal photograph, substantially edited with AI11
Frequently asked questions
Does an AI-generated voice need to be labelled?
Not every synthetic voice is automatically a . A realistic voice clone or a recording that sounds like a real person or event must be clearly disclosed as artificially generated or manipulated no later than first exposure. With a live voice assistant, people must also be able to tell from the start of the interaction that an AI is responding.
Is an AI notice in alt text or metadata enough?
Generally not for a required disclosure. The notice must be clear and perceivable to people; machine-readable marking alone does not replace it. Alt text remains important for accessibility. For audio, an audible notice is possible, but it should preferably be accompanied by text in the player.
Must the notice be permanently embedded in the image, audio or video?
The AI Act does not literally require this in every case. An equivalent interface overlay may suffice if it clearly appears on the content and is perceivable at first exposure. For exported or reshared files, an embedded notice is usually more robust because it travels with the file.
Who is responsible if the AI tool does not provide a marking?
The provider of the generative AI system is generally responsible for machine-readable marking. The publishing business remains responsible for adding any required clear disclosure. Missing metadata does not remove that duty.
Can I label voluntarily even if it is probably not required?
Yes. A factual voluntary label can build trust and reduce uncertainty. It does not replace checks of advertising claims, personality rights, trade marks or copyright.
Does the AI Act apply to small businesses and international freelancers?
The rules depend on the actor’s role and use, not on company size or citizenship. Activity from a third country may also be covered where the AI output is used in the EU. Enforcement must be proportionate, and the Act contains specific considerations for SMEs.
12
Official sources and related rules
Primary sources and official guidance used for this article. Reviewed and expanded on 11 August 2026.
- Regulation (EU) 2024/1689
- European Commission: Article 50 Q&A
- European Commission: Article 50 Guidelines
- European Commission: transparency quick facts
- European Commission: AI literacy Q&A
- European Commission: Transparency Code
- European Commission: labelling icons
- Unfair Commercial Practices Directive
- General Data Protection Regulation
- EU Copyright Directive
- EU Trade Mark Regulation
13
Does my AI content need special attention?
Free 2-minute check
A few short questions provide an initial, non-binding orientation under Article 50. This check neither transmits nor stores your answers.
No email · No registration · Runs only in your browser
General information, not legal advice. Whether a specific publication must be disclosed always depends on its content, context and intended audience. Assessment based on Article 50 and the European Commission guidelines.