Do I need to label that? A practical AI content guide for business teams
Ready 4 AI Act · August 3, 2026 · Last updated August 10, 2026
From 2 August 2026, the EU AI Act's transparency rules apply to several common ways businesses use AI.
The good news: not every AI-generated image, video, email or social post needs an AI label.
For most business teams, the decision comes down to a few practical questions:
- Does an AI-generated image, video or voice look real enough to be mistaken for authentic content?
- Is AI-generated text being published to inform the public about an important public-interest topic?
- Is someone directly interacting with an AI system, such as a chatbot or voice agent?
- Are you using emotion recognition or biometric categorisation on people?
This guide applies those questions to everyday business use cases.
Important: This article focuses mainly on businesses using AI tools. AI tool providers have additional technical obligations, including machine-readable marking of certain AI-generated content.
First: three simple rules to remember
Rule 1: Real-looking AI images, video and audio
If AI-generated or manipulated content could reasonably be mistaken for authentic content, you should disclose that it was generated or manipulated by AI.
Think about realistic content showing:
- people
- products or other objects
- places
- organisations or other entities
- events
The key question is simple:
Could someone reasonably think this is a real photo, recording or piece of footage?
If yes, disclose.
If it is clearly an illustration, animation, abstract graphic or obviously synthetic content, a visible AI label will usually not be required under the deepfake rule.
Rule 2: AI-written public-interest content
Most everyday AI-written business text does not need an AI label.
The specific text rule matters when AI-generated or AI-manipulated text is published to inform the public about a matter of public interest.
Examples can include:
- politics and elections
- public health
- public safety
- law and public administration
- consumer safety
- environmental issues
- important economic or financial developments
- scientific or cultural developments that are part of public debate
There is an important exception:
If a knowledgeable person genuinely reviews the substance of the text and your organisation takes editorial responsibility for publishing it, an AI label is not required under this rule.
A spellcheck, grammar check or quick skim is not enough.
Rule 3: People interacting directly with AI
If someone is having a genuine conversation with an AI system, they should know it is AI from the start.
Examples include:
- customer-service chatbots
- AI voice agents
- interactive AI avatars
- conversational AI assistants
There is an exception where it is already obvious to a normal user that they are interacting with AI.
As a practical business rule, however:
If there is any realistic chance that someone could think they are dealing with a human, disclose that it is AI.
Social media content
Topic: AI-generated social media images
What applies:
A visible disclosure is needed when the image looks like an authentic photograph and could reasonably be mistaken for something real.
Why:
Realistic AI-generated images can fall under the AI Act's deepfake transparency rule.
Examples:
- Clearly illustrated campaign graphic - no AI label normally needed
- Abstract AI background - no AI label normally needed
- Photorealistic AI image of a fictional employee - label it if it could look like a genuine photograph
- AI-generated photo of a fake office opening - label it if viewers could believe the event really happened
Topic: AI-written LinkedIn, Instagram or Facebook posts
What applies:
Normal marketing posts, product updates and company news will usually not need an AI label just because AI helped write them.
Why:
The AI Act's text-labelling rule is focused on published information about matters of public interest.
If the post covers a public-interest topic, substantive human review and editorial responsibility can remove the labelling requirement.
Practical action:
If you publish AI-assisted posts about politics, regulation, public health, safety, major economic developments or similar subjects, make sure a knowledgeable person reviews the substance before publishing.
Topic: AI-generated Reels, TikToks and Shorts
What applies:
Label realistic AI-generated or manipulated video when viewers could mistake it for genuine footage.
Why:
The deepfake rule applies to realistic AI-generated or manipulated video, not only still images.
Examples:
- Animated explainer - no visible AI label normally needed
- Clearly fictional 3D scene - normally no label
- AI-generated footage of people attending an event that never happened - label it
- Realistic AI video of a person recommending your product - label it if it could appear authentic
Topic: AI avatars in social videos
What applies:
- Pre-recorded and clearly animated avatar - normally no disclosure
- Pre-recorded photorealistic AI presenter - disclose if it could be mistaken for authentic footage of a person
- Live interactive AI avatar - tell users they are interacting with AI unless this is already obvious
Why:
A pre-recorded presenter is mainly a content question. A live avatar that responds to people is an interactive AI system.
E-commerce
Topic: Website chatbot
What applies:
Visitors should know from the beginning that they are interacting with AI, unless that is already obvious.
Why:
The AI Act has a specific transparency rule for AI systems designed for direct two-way interaction with people.
Practical action:
Check that your chatbot clearly identifies itself as AI before or at the start of the conversation.
If you use a third-party chatbot provider, verify that this disclosure is built into the customer experience.
Topic: AI-generated FAQ and help-centre content
What applies:
Delivery information, return policies, sizing guides and normal product FAQs usually do not need an AI label.
Why:
They are generally not published to inform the public about matters of public interest.
Practical action:
Still have a human review important customer information. The main risk here is usually incorrect or misleading information rather than Article 50 labelling.
Topic: AI-generated product images
What applies:
A realistic synthetic product image may need disclosure if customers could reasonably believe it is an authentic photograph.
Why:
The deepfake definition can cover realistic AI-generated images of objects, including products.
Examples:
- Clearly stylised product concept - normally no AI label
- Abstract product visual - normally no AI label
- Photorealistic AI product photo presented like a real studio photograph - consider a visible AI disclosure
- AI image showing product features that the real product does not have - high risk of misleading customers, regardless of the AI Act labelling question
Practical action:
Ask two questions:
- Could a customer think this is a real photograph?
- Could the image create a false impression of the product?
If either answer is yes, treat it carefully.
Marketing
Topic: AI-generated models in advertising
What applies:
If an AI-generated person looks like a real human photographed or filmed for the campaign, disclose that the content is AI-generated.
Why:
A person does not have to be a specific real individual before realistic synthetic content can fall within the deepfake rules.
A completely fictional AI model can still look authentic enough to require disclosure.
Example:
A photorealistic AI fashion model that looks indistinguishable from a real campaign photo - disclose.
Topic: AI-generated backgrounds, locations and scenes
What applies:
Clearly fictional or stylised scenes usually do not need disclosure. Realistic synthetic locations or scenes may.
Why:
The rule looks at whether the content could falsely appear authentic.
Examples:
- Illustrated future city - normally no label
- Abstract background - normally no label
- Photorealistic image labelled as a concept design - lower risk because the context makes its nature clear
- AI-generated image presented as "our new Copenhagen office" when that office does not exist - disclose and do not present it misleadingly
Topic: AI-written adverts, landing pages and email campaigns
What applies:
Normal commercial copy usually does not require an AI label.
Why:
The public-interest text rule is not a general rule requiring businesses to label all AI-written text.
Watch out when:
Your campaign is publishing information about public health, safety, politics, regulation, important financial developments or another public-interest matter.
In those cases, ensure meaningful human review and clear editorial responsibility before publication.
Topic: AI-generated voiceovers
What applies:
A clearly synthetic voice will normally not need a deepfake disclosure. A cloned or highly realistic voice that could be mistaken for an authentic recording should be disclosed.
Why:
The deepfake rules apply to audio as well as images and video.
Examples:
- Clearly robotic text-to-speech voice - normally no visible label
- Realistic synthetic narrator presented simply as narration - assess the context
- Cloned voice of a real executive - disclose
- AI voice made to sound like a real customer testimonial - disclose and check other advertising and consumer-protection rules
Sales
Topic: AI-generated visuals in a sales deck
What applies:
A realistic AI-generated image can need disclosure even when it is shown privately to one prospect.
Why:
The deepfake disclosure rule is not limited to public social media posts.
Practical action:
If a realistic image in your deck is synthetic, label it on the slide or next to the image so the prospect knows at first exposure.
Topic: AI-drafted sales emails and proposals
What applies:
A normal one-to-one sales email or proposal does not need an AI label under the public-interest text rule.
Why:
It is not normally being published to inform the public.
Practical action:
Use human review for accuracy, pricing, commitments and claims, but an AI disclosure is generally not required merely because AI helped draft the message.
Topic: AI sales chatbot or voice agent
What applies:
Tell prospects they are interacting with AI unless that is already obvious.
Why:
This is direct interaction between a person and an AI system.
Practical action:
Make the disclosure clear at the start, not several messages or minutes into the conversation.
Topic: AI meeting note-taker
What applies:
A passive AI note-taker or internal summariser will generally not trigger Article 50 content-labelling requirements.
Why:
It is normally working in the background rather than publishing synthetic content to the public or directly pretending to be a human participant.
Watch out:
Recording and processing meeting participants can raise separate privacy, confidentiality and GDPR questions.
Events
Topic: AI-generated event graphics and banners
What applies:
Stylised graphics normally do not need an AI label. Photorealistic synthetic images may.
Why:
The question is whether the content could reasonably be mistaken for an authentic photograph.
Examples:
- Illustrated conference banner - normally no label
- Abstract AI background - normally no label
- AI-generated photo showing a packed venue before the event has taken place - disclose if it could appear to be real event photography
Topic: AI-written speaker bios, agendas and event copy
What applies:
Normal event copy usually does not need AI labelling.
Why:
Most event information is not covered by the public-interest text rule.
Practical action:
Always review factual information about speakers, dates, credentials and agendas before publishing.
If the event content itself covers important public-interest topics and AI drafts substantive public information, use meaningful human review and editorial responsibility.
Topic: AI virtual host or emcee
What applies:
- Live interactive AI host - tell attendees it is AI unless this is already obvious
- Pre-recorded, clearly animated host - normally no disclosure
- Pre-recorded photorealistic AI presenter - disclose if it could be mistaken for authentic footage of a real or plausible person
Why:
Live interaction and pre-recorded synthetic content fall under different transparency tests.
Topic: AI-enhanced event recap videos
What applies:
Normal editing of real footage usually does not create a labelling obligation. Fabricated realistic scenes can.
Why:
Cuts, colour correction and ordinary post-production normally do not turn genuine footage into a deepfake.
Examples:
- Colour correction - no AI disclosure normally needed
- Automatic subtitles - no deepfake disclosure
- Removing background noise - no deepfake disclosure
- Adding realistic AI-generated audience scenes that never happened - disclose
- Creating a fake clip of a speaker saying something they never said - disclose
Blogs and articles
Topic: AI-assisted blog text
What applies:
An article does not need an AI label simply because AI helped draft it.
If the article informs the public about a matter of public interest, the text rule may apply.
Why:
Public-interest AI-generated text can require disclosure, but there is an important exception for genuine human review and editorial responsibility.
Practical action:
For articles about regulation, politics, health, safety, economic developments or similar topics:
- Have someone with suitable knowledge review the substance.
- Check facts and sources.
- Correct errors and misleading statements.
- Make sure a person or organisation takes responsibility for the final publication.
A quick grammar check is not enough.
Topic: AI-generated blog cover images
What applies:
Stylised graphics and illustrations normally do not require a visible AI label.
Photorealistic synthetic images may.
Why:
The deepfake rule is about content that could falsely appear authentic.
Examples:
- Abstract AI illustration - normally no disclosure
- Stylised legal-tech graphic - normally no disclosure
- Photorealistic image of a fictional CEO speaking at a conference - disclose if readers could think it is a real photograph
Topic: Content published before 2 August 2026
What applies:
Content generated before 2 August 2026 does not need to be relabelled retrospectively under Article 50.
Why:
The Commission has clarified that the new transparency rules do not require businesses to go back and relabel older content.
Practical action:
You can still add voluntary disclosures where useful, especially for realistic synthetic content that could otherwise confuse people.
Topic: AI-generated charts and infographics
What applies:
A normal chart or infographic does not usually need a deepfake label.
Why:
Readers normally understand charts and infographics as designed information graphics rather than authentic photographs or recordings.
Watch out:
If the graphic contains realistic synthetic photos or scenes, apply the image rule.
If AI-generated explanatory text in the graphic informs the public on a public-interest matter, apply the public-interest text rule to that text.
And regardless of labelling, always verify the underlying data.
Special case: emotion recognition and biometric categorisation
Topic: Using AI to analyse people's emotions or biometric characteristics
What applies:
People need to be informed when your business uses an AI emotion-recognition or biometric-categorisation system on them, where the use is permitted under the AI Act.
Why:
Article 50 contains a specific transparency obligation for these systems.
Examples can include:
- AI analysing facial expressions to infer emotions
- AI analysing voice patterns to infer emotional state
- AI categorising people using biometric characteristics
Practical action:
Do not treat these like ordinary analytics tools. Check the AI Act rules for the use case before deployment and make sure affected people receive the required information.
Some uses may also be restricted or prohibited under other parts of the AI Act.
What about watermarks and hidden metadata?
Topic: Machine-readable AI marking
What applies:
If your business is simply using a third-party generative AI tool, you will normally not be responsible for creating the tool's technical machine-readable marking.
Why:
That technical obligation sits with the provider of the generative AI system.
But that does not replace your own visible disclosure when you publish a qualifying deepfake.
Practical example:
A generated image might already contain invisible machine-readable AI information.
If you publish that image and it looks like a genuine photograph, you may still need to tell people visibly that it is AI-generated.
2026 transition
Providers of generative AI systems that were already placed on the market before 2 August 2026 have until 2 December 2026 to meet the Article 50(2) machine-readable marking requirement.
This transition does not generally postpone the business user's separate obligation to disclose qualifying deepfakes or public-interest AI text.
The quick business checklist
Before publishing or deploying AI-generated content, ask:
1. Does it look or sound real?
Could someone reasonably mistake the AI-generated or manipulated image, video or audio for authentic content?
Yes: Add a clear disclosure.
No: A deepfake disclosure is normally not required.
2. Is it AI-generated text about a public-interest topic?
Is the text being published to inform the public about politics, public health, safety, law, consumer safety, important economic developments or another matter of public debate?
No: The Article 50 public-interest text label normally does not apply.
Yes: Ask the next question.
3. Has a knowledgeable human genuinely reviewed it?
Did someone review the substance, fact-check it where needed and take responsibility for publication?
Yes: An AI label is normally not required under the public-interest text rule.
No: Clearly disclose that the text was AI-generated or manipulated.
4. Is someone directly talking to AI?
Is a customer, employee, prospect or attendee having a genuine two-way conversation with an AI system?
Yes: Make sure they know it is AI from the start, unless this is already obvious.
No: The direct-interaction disclosure rule does not apply.
5. Are you analysing emotions or biometric characteristics?
Yes: Stop and check the specific AI Act requirements before deployment. People may need to be informed, and some uses face additional restrictions.
A practical rule of thumb
If you remember only one thing, remember this:
The EU AI Act is not asking businesses to label everything made with AI. It is mainly asking for transparency where people could be misled about what is real, where important public information is published without meaningful human oversight, or where people are directly dealing with AI.
When in doubt, a simple disclosure such as "AI-generated image", "AI-generated video" or "You are chatting with an AI assistant" is often the clearest approach.
This is general practical guidance, not legal advice. The right answer can depend on the content, audience and way the AI system is used.
If you want to understand which AI Act obligations apply across your organisation, you can run the free Ready 4 AI Act assessment.
Official sources
Reviewed by Ready 4 AI Act EU - Editorial team. This article is journalistic information, not legal advice.
Not sure where you stand?
Take the free readiness assessment and get a personalised report.
Start free assessment