Will that disclaimer soon have to appear under every blog post ChatGPT helped write? Does every product image created with artificial intelligence need to be labeled? And does every social media graphic need a note about the AI tool behind it?
The short answer is: no. The EU AI Act does not require a blanket, visible label on every piece of content that AI was involved in creating.
Since August 2, 2026, however, the transparency duties of Article 50 of the European AI Act have applied. They cover, among other things, AI-generated and manipulated content, deepfakes, certain publications on matters of public interest, and systems that interact directly with people.
For businesses, agencies, bloggers, content creators and the self-employed, it's therefore crucial to understand when a label is actually required — and when it isn't. That's exactly what this article looks at.
The most important point up front: a blog post drafted with ChatGPT and then reviewed by a subject-matter expert is legally a very different thing from an AI-generated deepfake video that makes a real person appear to say something they never said.
What does Article 50 of the EU AI Act actually regulate?
Article 50 is titled "Transparency obligations for providers and deployers of certain AI systems." The AI Act distinguishes mainly between two roles here:
Providers develop or market an AI system, or place it on the market under the relevant conditions. Deployers use an AI system under their own responsibility. That can include businesses, agencies, self-employed professionals or public authorities that use generative AI professionally.
Anyone using ChatGPT, Claude, Gemini or another ready-made AI system in their business is therefore not automatically a provider of the underlying AI model themselves. For ordinary users, the deployer role is the relevant one.
The transparency rules can be simplified into four areas:
- Under certain conditions, people must be able to recognize that they are interacting with an AI system.
- Providers of generative AI systems must mark certain AI outputs in a machine-readable format.
- Deployers must disclose deepfakes.
- Certain AI-generated or manipulated text on matters of public interest must be labeled.
The European Commission likewise summarizes these as the core transparency duties of Article 50.
Has the labeling requirement applied since August 2026?
Yes. Article 50 has applied since August 2, 2026. Businesses should no longer rely on the AI Act's long transition period.
There is, however, a limited special rule: generative AI systems that were already placed on the market before August 2, 2026 get a transition period until December 2, 2026 for the technical, machine-readable marking required under Article 50(2). This change was introduced through the Digital Omnibus.
Important: this is not a general grace period until December 2026. It specifically concerns providers of already-existing generative AI systems and their technical marking duties. The transparency duties for deployers — for example around deepfakes or certain publications — have generally applied since August 2, 2026.
Does every AI-generated text need to be labeled?
No. This is exactly where one of the biggest misunderstandings about the AI Act currently comes from.
Article 50(4) does not require that every text created commercially with ChatGPT, Claude or Gemini automatically carry the label "AI-generated." The specific disclosure duty concerns text that was generated or manipulated with AI and is published to inform the public on matters of public interest.
That can include topics such as politics and democratic processes, public administration, justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, economic and financial developments, scientific developments, and political or cultural topics that are the subject of public debate.
An automatically generated news piece about a health crisis can therefore be judged very differently from an AI-assisted product description for screws in an online shop.
What about blog posts?
For blog posts too, the answer isn't automatically "AI was used, so a label is required." What matters is the content and the editorial process.
A blog post can certainly touch on matters of public interest. For such text, though, Article 50 contains an important exception: disclosure is not required if the AI-generated content has gone through a process of human review or editorial control and a natural or legal person takes on editorial responsibility for the publication.
That's highly relevant for professional websites and editorial teams. A possible process might look like this: AI drafts the content, a knowledgeable person checks claims and sources, errors are corrected, content is added or removed, the final version is deliberately approved, and the deployer takes responsibility for the published result.
In that case, there's a substantial difference from a platform that automatically generates hundreds of AI texts and publishes them without any editorial control.
Clicking "publish" is not the same as human review
The European Commission has since clarified in more detail what it means by human review: a deliberate, substantive check by people who have sufficient knowledge and professional judgment on the relevant topic.
Editorial control, too, means more than fixing spelling and formatting. The responsible person must genuinely be able to decide whether content gets published, changed or rejected — including fact-checking and assessing the sources used. A purely superficial correction of grammar or spelling is not enough under the guidelines.
That has real practical consequences. A sound company process therefore isn't "let ChatGPT write it → skim it → publish," but rather:
Let AI assist → check the facts → verify the sources → revise if needed → responsible sign-off → publish.
Example: does an AI-assisted company blog need a label?
Say an IT company publishes an article titled "How the new EU cybersecurity directive works." AI produces a first draft. A staff member then checks the legal basis, verifies primary sources, corrects claims, adds their own explanations, and takes editorial responsibility.
In that case, the exception for human-reviewed, editorially accountable text can apply. It looks different if an automated system independently generates an article on a matter of public interest and that article is published without any genuine substantive review. Then the labeling duty under Article 50(4) can become relevant.
Do AI-generated images need to be labeled?
Here too: not every AI-generated image must visibly carry the label "AI-generated" simply because of that. You need to distinguish between the technical marking done by the provider of a generative system and the deployer's disclosure duty.
Providers of AI systems that generate synthetic images, audio, video or text must generally ensure that these outputs are marked in a machine-readable format and can be recognized as artificially generated or manipulated. That technical duty is primarily directed at the provider of the system.
A small business therefore doesn't automatically have to build its own invisible watermarking system for every marketing image created with an image generator. For businesses acting as deployers, the more relevant question is: is this a deepfake?
What counts as a deepfake under the AI Act?
The term is often used far more broadly than the AI Act actually means it. In the context of Article 50, it concerns AI-generated or manipulated images, audio or video content that resembles existing people, objects, places, entities or events and could falsely appear to be authentic or true.
An obviously illustrated, fantastical image of a robot is therefore not automatically a deepfake. It would be different if, for example, AI produces a photorealistic video of an actual CEO and has them say things they never said. Or if a synthetic recording shows a real building apparently at an event that never happened. It's exactly in scenarios like these that transparency becomes especially important.
Deepfakes must be disclosed
Anyone who, as a deployer, uses an AI system to generate or manipulate image, audio or video content that constitutes a deepfake must disclose that the content was artificially generated or altered. That disclosure must generally be made no later than the first exposure to the person concerned, and it must be clear and distinguishable.
A purely invisible technical marker is not sufficient to meet this disclosure duty. The Commission points, for example, to visible or audible labels, such as:
"This image was created using artificial intelligence."
Or, for manipulated content:
"This video contains AI-generated or AI-manipulated elements."
The disclosure should sit right where the content is perceived — not hidden away on some separate subpage.
Are there exceptions for art, satire and fiction?
Yes. The AI Act explicitly accounts for artistic, creative, satirical, fictional and comparable works. For corresponding deepfake content, disclosure should be made in a way that doesn't unduly interfere with the display or enjoyment of the work. A film scene with AI-generated effects doesn't have to be plastered with a giant warning in the middle of the frame. Transparency still needs to be established appropriately, though.
Can I label AI content voluntarily?
Of course. The AI Act sets minimum requirements. Businesses are free to be more transparent than that. Especially for AI-generated images, voices or videos, voluntary labeling can even build trust.
That said, it's worth distinguishing three things: legally required labeling, voluntary transparency, and the technical provenance marking applied by the AI system itself. These three areas get mixed together quite often in discussions.
The European Commission has since published its own symbols for labeling AI-generated or manipulated content. Using them can help businesses present this consistently. The EU also recommends using such symbols prominently and accessibly, and phrasing accompanying text as clearly as possible. That doesn't mean every AI image on the internet must carry this symbol — it's a tool to help with practical implementation.
What does "machine-readable" actually mean?
One of the most important — and simultaneously least visible — changes under the AI Act concerns technical provenance detection. Article 50(2) generally requires providers of generative AI systems to mark artificially generated or manipulated audio, image, video and text content in a machine-readable format so it can be recognized as such.
The methods used should be effective, interoperable, robust and reliable, as far as technically feasible. Exactly how this provenance information is implemented technically depends on the content type and the state of the art — for example through embedded provenance data or other technical mechanisms that can be evaluated automatically.
Article 50(2) also has exceptions: the duty to apply machine-readable marking doesn't apply to the same extent where the AI system merely performs an assistive function in a standard editing process, or doesn't substantially alter the input data or its meaning. The Commission's current guidelines also list various special cases and exclude certain purely technical or machine outputs from the relevant scope.
That's also why the simple rule "AI was involved somewhere, so it has to say AI" would be legally too sweeping.
What applies to chatbots and AI agents?
Article 50 isn't only about content. Providers of AI systems that interact directly with people must design those systems so that users are informed they are communicating with an AI system — unless that's already obvious from the context.
A business running an AI chatbot on its website should therefore factor this transparency duty into choosing and configuring the system. A simple disclosure might read: "You're chatting with our AI assistant." People shouldn't be deliberately left with the impression that they're speaking to a real support agent when, in fact, only an AI is responding.
Practical examples: label or not?
| Example | Visible labeling under Art. 50 typically required? |
|---|---|
| ChatGPT helps draft a product description | Not simply because AI was used |
| AI drafts a blog post that is reviewed and editorially owned by a human | For text on matters of public interest, the human-review exception can apply |
| AI automatically writes news on a public-interest topic without human review | Labeling can be required |
| Obviously artificial illustration of a fantasy robot | Not automatically a deepfake requiring labeling |
| Photorealistic AI video of a real politician making a fabricated statement | Deepfake — disclosure required |
| AI voice convincingly imitates a real person | Check the deepfake rules; disclosure can be required |
| AI merely corrects spelling in a text | No blanket labeling duty |
| AI chatbot talks with customers on a website | Ensure transparency about the AI interaction |
| AI-generated satirical film with realistically recreated people | Special disclosure rule for creative/satirical works |
The specific individual case always remains decisive.
Is a general notice in the legal notice page enough?
For a legally required deepfake disclosure, a notice buried somewhere in a legal notice or imprint page would typically defeat the purpose of the rule.
Article 50(5) requires that relevant information be provided clearly and distinguishably no later than the first interaction or exposure. Applicable accessibility requirements must also be taken into account. In practice, that strongly favors placing disclosures directly at the content in question: for a video, for example, under or on the video itself; for audio content, as a visible or, where relevant, audible notice; for affected text, directly alongside the article.
Does old AI content need to be labeled retroactively?
Under the European Commission's current explanations, generally not. AI-generated or manipulated content that was already created and published before August 2, 2026 does not need to be retroactively labeled because of Article 50. The Commission does, however, recommend voluntary labeling where it's sensible and feasible.
Businesses therefore don't need to automatically comb through their entire website archive looking for generative AI that may have been used at some point.
Does every business now need an AI labeling policy?
A hundred-page compliance policy probably isn't necessary for a small business. A clearly defined process is still worthwhile, though. Businesses should define at least the following:
- Which generative AI systems are being used?
- What content are they used for?
- Are real people, voices, places or events depicted synthetically?
- Does the text concern a matter of public interest?
- Does genuine human review take place?
- Who takes on editorial responsibility?
- Where does a visible disclosure need to be placed, if any?
- Does your processing preserve the AI system's technical provenance markers?
These questions fit neatly into an existing internal AI policy. For a template with a traffic-light model, see AI Usage Policy for Companies: What Employees May and May Not Do (Read article).
Document your editorial review
Businesses that regularly use AI-generated drafts for blog posts, technical content or corporate communication should not only carry out their own approval process in practice, but also define it organizationally. For example:
AI draft → source check → subject-matter review → editorial revision → sign-off by the responsible person → publication.
That doesn't need to create complicated bureaucracy. A traceable editorial process alone can already be far better than an undefined "someone looks it over first." After all, the exception for certain AI-generated text depends precisely on genuine human review or editorial control actually taking place, with someone taking responsibility for the publication.
How much of this is happening without any control at all in your business is often only clear after an honest inventory — more on that in Shadow AI at Work: When Employees Secretly Use ChatGPT (Read article).
Matching product · German-language edition
KI ohne Bauchschmerzen
This German-language practical guide helps small businesses introduce AI in a controlled way, with clear internal rules, risk assessment, 50 prompt templates, a 21-day plan, and a sensible sign-off process for AI-assisted publications.
The voluntary EU code of practice
On June 10, 2026, the final "Code of Practice on Transparency of AI-Generated Content" was also published. The code is voluntary but is meant to help businesses and AI providers put the legal transparency duties into practice.
The legal requirements under Article 50 itself, however, are not voluntary. Signatories can use the measures described in the code to demonstrate compliance with the transparency requirements. Businesses that don't sign it may use other suitable methods, but should still be able to explain and document their approach.
What penalties are possible?
The AI Act shouldn't be viewed solely through the lens of possible fines. Still, the sanctions it provides for are substantial. For breaches of various AI Act obligations, fines of up to €15 million or up to 3% of worldwide annual turnover can apply.
For SMEs, the AI Act's specific proportionality rules apply. The nature, severity, duration and circumstances of a breach also play a role. For smaller businesses, that doesn't mean a forgotten label triggers a multi-million-euro fine overnight. But it does mean: the transparency requirements are now genuine, applicable law and shouldn't be treated as a theoretical future rule anymore.
For a full overview of roles, duties and the fine framework, see EU AI Act for Small Businesses: What SMEs Need to Know Now (Read article).
Checklist: does my AI content need a label?
Before publishing AI-assisted content, these questions help:
- Is it an image, audio or video? Check whether it constitutes a deepfake within the meaning of the AI Act.
- Does it convincingly depict a real person, place or event? The likelihood that disclosure is required rises significantly.
- Is it text? Check whether it's being published to inform the public on a matter of public interest.
- Was this text genuinely reviewed on substance and editorially? A quick grammar check isn't enough.
- Is there a natural or legal person taking on editorial responsibility? That's decisive for the exception that applies to certain text.
- Is the label visible or otherwise perceivable at first exposure? A notice hidden somewhere on the website isn't the same thing.
- Are you using a chatbot or AI agent with customers? Check whether it's sufficiently transparent that the person is interacting with an AI system.
- Could your processing strip out technical AI provenance markers? Your publishing workflow should account for this too.
Frequently asked questions about labeling AI-generated content
Do I have to label every ChatGPT text?
No. The AI Act does not require a blanket label on every text created with ChatGPT. Article 50(4) becomes relevant specifically for published AI-generated or manipulated text on matters of public interest. An important exception applies to content that underwent human review under editorial responsibility.
Does an AI-assisted blog post need to be labeled?
Not automatically. For topics of public interest, you should check whether genuine human review or editorial control took place and whether a person or organization has taken on editorial responsibility.
Does every AI-generated image need to be labeled?
Not automatically and not always visibly. Providers of generative systems have technical marking duties. Deployers must in particular disclose AI-generated or manipulated images if they constitute a deepfake within the meaning of the AI Act.
Does an AI-generated video need to be labeled?
If it constitutes a deepfake, generally yes. Special rules for appropriate disclosure apply to works that are evidently artistic, creative, satirical or fictional.
Is a watermark in the metadata enough?
Not for a deepfake disclosure duty. The Commission makes clear that deployers cannot rely solely on invisible machine-readable markers for this. The information must be perceivable by people.
Is a quick edit of AI text enough to count as review?
Not necessarily. A purely superficial check of spelling, grammar or formatting does not count as sufficient human or editorial review under the guidelines.
Does content from before August 2026 need to be labeled retroactively?
Under the Commission's current guidance, generally no. Voluntary retroactive labeling is nonetheless encouraged.
Does a website need to state that a chatbot uses AI?
For systems that interact directly with people, it must generally be ensured that users can recognize they are communicating with an AI system — unless this is already obvious from the context.
Bottom line: not "label everything with AI," but tell the real cases apart
The EU AI Act's new transparency duties are far more nuanced than some headlines suggest. Article 50 has applied since August 2, 2026. Even so, that doesn't create a general duty to stamp "AI-generated" under every paragraph ChatGPT helped write and every marketing image created with AI.
What actually matters most: deepfakes, certain text on matters of public interest, direct interactions with AI systems, and the technical provenance marking of generative systems. For text, genuine human editorial control plays an additional, decisive role.
That leaves businesses with a fairly practical strategy: let AI prepare content. Have humans review relevant publications on substance, take responsibility, and label transparently wherever the AI Act actually requires it. That's not just good for compliance — it also keeps businesses from reflexively slapping a warning label on every AI-assisted sentence out of fear of the AI Act.
Sources and legal status
Editorial status: October 15, 2026. The main sources are Regulation (EU) 2024/1689 – Artificial Intelligence Act, in particular Article 50, the amendments introduced by the Digital Omnibus on AI, the European Commission's guidelines on Article 50 and the Code of Practice, as well as the AI Act Explorer: Article 50 and AI Act Explorer: Article 99. Because further guidance keeps being issued, businesses should check the current status regularly.
Note: This article is for general information only and does not replace individual legal advice.