Does AI-Generated Content Have to Be Labelled Under the EU AI Act?
The EU now requires some AI-generated content to be labelled. The difficult part is knowing whether yours is one of the cases covered.
You use AI to help draft a blog post. A designer creates an image with a generative tool. A communications team publishes an AI-written summary of a new government policy. A parody account posts a synthetic video of a politician.
Do all four pieces of content need the same “AI-generated” label?
No. The EU AI Act does not create one universal badge that must be attached to everything touched by AI. Article 50 creates different transparency duties depending on who is responsible, what kind of content is involved, how much the AI changed it and what the content is being used for.
The practical question behind AI content labelling is, therefore, not simply whether AI was involved. It is which Article 50 duty applies to the content and the person or organization responsible for it.
That matters because a lot of online discussion collapses two separate ideas into one: machine-readable marking by the AI provider and clear disclosure to people by the organization using the AI. They overlap, but they are not the same legal obligation.
This guide explains the current rules that have applied since 2 August 2026, with a practical focus on generated text, images, audio and video. It is general information rather than legal advice, but it should make the first question much easier to answer: does this particular AI-assisted content actually fall into one of the cases that requires marking or disclosure?
Does AI-Generated Content Have to Be Labelled Under the EU AI Act?
Some AI-generated content must be marked or disclosed under the EU AI Act, but not every AI-assisted output needs a visible label.
Article 50 separates the transparency framework into several situations. Providers of certain AI systems have duties to make generated or manipulated content detectable in machine-readable form. Deployers, meaning organizations or people using AI systems in a professional context, have separate duties to inform people in particular cases, including deepfakes and certain AI-generated text published to inform the public on matters of public interest.
The European Commission’s current Article 50 quick facts summarizes the framework as four transparency cases: direct interaction with AI, synthetic-content marking, emotion recognition or biometric categorization, and exposure to deepfakes or certain public-interest text.
| Situation | Main responsibility | What the rule is trying to achieve |
|---|---|---|
| Person interacts directly with an AI system | Provider | Make it clear that the person is interacting with AI, unless that is already obvious. |
| AI system generates synthetic text, image, audio or video | Provider | Make the output machine-detectable as AI-generated or manipulated, subject to exceptions. |
| Professional user publishes a deepfake | Deployer | Disclose that the content has been artificially generated or manipulated. |
| Professional user publishes certain AI-generated public-interest text | Deployer | Disclose AI generation unless the human-review/editorial-responsibility exception applies. |
The wrong conclusion is that every use of ChatGPT, Claude, Gemini or another generative system must produce the same visible warning. The law first asks which duty applies.
AI Marking and AI Labelling Are Not the Same Thing
The easiest way to understand Article 50 is to split it into two layers.
Layer 1: machine-readable marking
Article 50(2) applies to providers of AI systems that generate synthetic audio, image, video or text. Those providers must ensure that outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated, as far as technically feasible.
This does not necessarily mean a visible caption under the content. Machine-readable signals can include technical provenance or watermarking methods that software can detect. Vappingo’s guides to AI text watermarking, SynthID and Content Credentials and C2PA explain the main technologies that sit behind this idea.
Layer 2: human-facing disclosure
Article 50(4) applies to deployers in specific publication contexts. A deployer may need to tell people clearly that a deepfake or qualifying public-interest text was generated or manipulated by AI.
This is why a statement such as “the EU requires AI-generated content to be labelled” is incomplete. Sometimes the rule is about an invisible technical signal. Sometimes it is about a clear disclosure to the audience. Sometimes the content falls outside the relevant duty altogether.
Provider or Deployer? Start With Who Is Responsible
Many misunderstandings start by treating the AI company and the person using the AI as if they have the same job.
Under the AI Act, a provider is the person or organization that develops an AI system, or has it developed, and places it on the EU market or puts it into service under its own name or trademark. A deployer is the person or organization using an AI system under its authority, except where the AI is being used in a purely personal, non-professional activity.
The Commission’s current Article 50 FAQ makes this division explicit because the transparency duties follow the role.
| Example | Likely role | Relevant transparency question |
|---|---|---|
| Company supplying a generative AI product | Provider | Does the system mark synthetic output and make AI interaction clear? |
| Newsroom using that product to create public-facing content | Deployer | Does the published content fall into a disclosure category? |
| University communications office using AI for public information | Deployer | Is the final publication public-interest text, and was there human review? |
| Individual using AI privately for personal brainstorming | Generally outside the deployer definition for that personal activity | Article 50 deployer disclosure is usually not the main issue. |
That last row is particularly important for ordinary users. The AI Act’s transparency regime is not written as a universal personal-labeling rule for every private interaction with a chatbot.
What AI Providers Must Mark in Machine-Readable Form
Article 50(2) covers AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text content. The provider must make the output machine-readable and detectable as artificially generated or manipulated.
The law does not mandate one single technical method. Instead, it says the solution should be effective, interoperable, robust and reliable as far as technically feasible, taking account of the content type, implementation costs and the generally acknowledged state of the art.
The Commission’s final Code of Practice on Transparency of AI-generated Content provides a voluntary route for providers and deployers to demonstrate compliance. The Commission and the AI Board concluded in July 2026 that the code adequately covers the relevant Article 50 obligations, although signing the code is voluntary and does not itself create conclusive proof of compliance.
That leaves room for different technical approaches. Watermarking, secured metadata, provenance information and detection mechanisms can all form part of the broader compliance picture.
The AI Act does not say every generated image needs the same visible EU badge
The provider-level duty is about machine-readable marking and detectability. The Commission has also introduced optional EU icons that deployers can use for human-facing disclosure, but use of those icons is not compulsory and an icon alone does not establish legal compliance.
That difference matters when assessing claims about “EU AI labels.” A visible label and a machine-readable provenance signal solve different problems.
When Deepfakes Need a Clear Label
The clearest deployer obligation applies to deepfakes.
Article 50(4) says that deployers using AI to generate or manipulate image, audio or video content that constitutes a deepfake must disclose that the content has been artificially generated or manipulated.
A deepfake is not simply any image created with generative AI. The Commission describes it as generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful to a person.
An AI image can be synthetic without being a deepfake
Example A: A clearly fantastical illustration of a dragon reading in a university library is AI-generated, but it does not falsely present a real event as authentic.
Example B: A realistic synthetic video making a real university vice-chancellor appear to announce a policy they never announced is much closer to the deepfake definition.
Creative, artistic, satirical and fictional works also receive more flexible treatment. Where deepfake content is part of an evidently artistic, creative, satirical, fictional or analogous work or program, disclosure still matters, but it can be made in an appropriate way that does not interfere with the display or enjoyment of the work.
The practical question is, therefore, not “Was generative AI used?” It is “Could a reasonable viewer mistake this manipulated representation for something authentic or truthful?”
When AI-Generated Text Needs a Disclosure
Text follows a narrower rule than many headlines suggest.
A deployer must disclose AI-generated or AI-manipulated text when it is published for the purpose of informing the public on matters of public interest. This can include contexts such as news, civic information, public-policy communication and other material intended to inform the public about matters that affect them.
That does not mean every AI-assisted paragraph on the internet automatically needs an Article 50 label.
A product description, a private email, a student’s ordinary coursework submission and an internal brainstorming note are not automatically transformed into “public-interest text” merely because AI was used to help write them. Other laws, platform rules, professional standards or academic-integrity policies may still apply, but they are separate questions.
This is one of the most useful corrections to the common claim that “the EU now requires all AI-written text to be labelled.” It does not.
Human Review Can Change the Rule for Public-Interest Text
Article 50 contains a major exception for public-interest text.
The disclosure duty does not apply where the AI-generated content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for publishing it.
Those two elements belong together. Human involvement is not simply a person clicking “approve.” The purpose of the exception is to distinguish automated publication from content for which a person or organization has exercised editorial control and accepted responsibility.
A university explains a change to national student-finance policy
Imagine a university communications team asks a generative AI system to draft a public article explaining a new national student-finance rule.
Scenario 1: The AI draft is published with only superficial checking. Because the purpose is to inform the public on a matter of public interest, the Article 50 disclosure duty may apply.
Scenario 2: The communications team checks the source legislation, corrects the analysis, rewrites the material as needed, applies its normal editorial process and the university assumes editorial responsibility for the publication. The specific public-interest-text disclosure exception can become relevant.
The difference is not that a human touched the file. The important point is accountability for the final publication.
Organizations should, therefore, avoid treating “human in the loop” as a decorative compliance phrase. If relying on the human-review exception, the review process and editorial responsibility need to be real.
Does AI-Assisted Editing Need to Be Marked?
Article 50(2) also contains an important provider-level exception for standard editing.
The machine-readable marking obligation does not apply to the extent that an AI system performs an assistive function for standard editing or does not substantially alter the input data supplied by the deployer or the meaning of that input.
The Commission’s 2026 guidelines include standard editing among the concepts and exceptions they clarify. This is particularly relevant to spelling correction, formatting assistance and other interventions that leave the underlying content and meaning substantially intact.
| Use of AI | Why the distinction matters |
|---|---|
| Correcting a typo or punctuation error | Closer to assistive standard editing and unlikely to represent substantial content generation. |
| Reformatting supplied text without changing meaning | Potentially within the standard-editing / non-substantial-alteration exception. |
| Rewriting a full section and adding new arguments | Much harder to characterize as standard editing because the output changes the substance of the input. |
| Generating an article from a short prompt | Clearly involves synthetic text generation rather than simple editing of supplied material. |
These examples are practical illustrations, not legal bright lines. Whether a particular workflow falls inside the exception depends on what the system actually does to the supplied material.
For writers and editors, the useful distinction is between assisting with an existing text and creating or materially transforming the content itself.
What Does This Mean for Blogs, Student Work and Business Content?
The quickest way to apply Article 50 is to stop asking whether AI was involved and ask what type of content and publication process you are dealing with.
| Scenario | Likely Article 50 position |
|---|---|
| Company publishes a realistic AI-generated video of its CEO appearing to say something they never said | Potential deepfake. Clear disclosure is likely required. |
| Blog uses AI to draft a general lifestyle article | Not automatically subject to the public-interest-text disclosure rule merely because AI assisted the writing. |
| News or public-policy page publishes AI-generated text with no human editorial review | Likely within the public-interest text rule if the purpose is to inform the public on a matter of public interest. |
| Public-interest article is substantively reviewed and published under human/editorial responsibility | The specific human-review and editorial-responsibility exception may apply. |
| Student privately uses AI to brainstorm an essay | Article 50 deployer labelling is generally not the main issue. University assessment rules are separate. |
| AI tool corrects spelling in supplied text without materially changing meaning | Standard-editing exception may be relevant to the provider marking duty. |
For students, this is an especially important separation. Whether the EU AI Act requires a public disclosure is not the same question as whether a university permits AI use in an assessment. Vappingo’s guide to how universities can investigate AI use explains the academic-integrity side separately.
For businesses, the risk is the opposite: assuming that a general “we use AI” policy somewhere on the website automatically deals with every Article 50 obligation. The law focuses on particular interactions and exposures, so disclosure may need to appear where the person encounters the relevant content.
Does the Rule Matter If You Are Outside the EU?
Potentially, yes.
The AI Act is not limited to companies physically established inside the European Union. The Commission’s Article 50 FAQ states that providers outside the EU can fall within the Act where the output of their AI system is used in the EU.
That matters for UK companies, US AI providers and international platforms serving EU users. Being established outside the EU does not automatically remove Article 50 from consideration.
The exact territorial analysis can become complex, particularly where providers, deployers, users and publication audiences are spread across several countries. Organizations with material EU exposure should, therefore, map the relevant workflow rather than assuming geography alone settles the question.
For a UK publisher, for example, a sensible first check is whether it is deploying an AI system in a way that exposes people in the EU to content covered by Article 50. If the answer may be yes, the next step is proper legal or compliance advice rather than an assumption that Brexit makes the AI Act irrelevant.
What Should Organizations Do Now?
The Article 50 transparency obligations have applied since 2 August 2026. The Commission has also published final guidelines and a voluntary Code of Practice to help providers and deployers demonstrate compliance.
A practical review does not need to begin with legal jargon. Begin with the content workflow.
Map the AI use before deciding what needs a label
- Identify the role: Are you providing the AI system or deploying someone else’s system?
- Classify the output: Text, image, audio, video, chatbot interaction, deepfake or something else?
- Check the purpose: Is text being published to inform the public on a matter of public interest?
- Document human review: If relying on editorial review, record who reviewed the content and who holds responsibility for publication.
- Separate technical marking from visible disclosure: Do not assume one automatically satisfies the other.
- Check accessibility: Required information must be clear, distinguishable and meet applicable accessibility requirements.
For providers, the Commission’s final Code of Practice is now an important benchmark for machine-readable marking and detection. For deployers, the Commission’s optional EU AI-content icons can help with disclosure design, but the Commission expressly notes that using an icon does not by itself prove legal compliance.
Frequently Asked Questions About AI Content Labelling
Does all AI-generated content have to be labelled under the EU AI Act?
No. Article 50 creates different obligations for providers and deployers. Providers of qualifying generative AI systems have machine-readable marking duties, while deployers have human-facing disclosure duties in specific cases such as deepfakes and certain public-interest text. Exceptions also apply.
Does AI-generated text need a visible label?
Not every AI-generated text does. The deployer disclosure rule specifically covers AI-generated or manipulated text published for the purpose of informing the public on matters of public interest. It also contains an exception where the content has undergone human review or editorial control and a person or organization holds editorial responsibility for publication.
Does ChatGPT content have to be labelled in the EU?
The answer depends on what the content is and how it is used. Article 50 does not impose one visible label on every passage produced with ChatGPT. Provider-level marking duties and deployer-level disclosure duties are different, and the visible text rule for deployers is tied to public-interest publications and the relevant exceptions.
Does an AI-generated image always need a label?
No. Providers have machine-readable marking duties for synthetic images, subject to the law’s limits and exceptions. The separate deployer disclosure duty applies to image, audio or video content that qualifies as a deepfake. A clearly fictional AI illustration is not automatically a deepfake.
Is a watermark the same as an AI label?
No. A watermark or provenance signal can be a machine-readable way to help detect AI-generated content. A visible disclosure is information presented to the person viewing the content. Article 50 uses both concepts for different responsibilities.
Do AI-assisted student essays need an EU AI label?
An ordinary student submission is not automatically covered by the deployer rule for text published to inform the public on matters of public interest. University academic-integrity and disclosure requirements are separate and can still require students to declare AI use even where Article 50 is not the relevant rule.
Does human editing remove the need to disclose AI use?
There is a specific exception for qualifying public-interest text where AI-generated content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication. That does not create a universal rule that any human edit removes every AI transparency obligation.
Are the EU AI-content icons compulsory?
No. The European Commission says use of its AI-content icons is optional. The underlying Article 50 disclosure obligations are mandatory where they apply, and using an icon does not by itself establish compliance.
When did the EU AI content labelling rules start?
Article 50’s transparency obligations started to apply on 2 August 2026. A limited transition applies to certain pre-existing systems for the provider marking-and-detection obligation, but the core transparency framework is now in force.
What are the penalties for breaking Article 50?
The Commission says fines for relevant infringements can reach up to €15 million or 3% of total worldwide annual turnover for the preceding financial year, with proportionality considerations for smaller organizations. The competent authority and penalty analysis depend on the circumstances.
The Label Question Starts With the Type of Obligation
The EU AI Act does require greater transparency around AI-generated and manipulated content. It does not do so through one universal “AI-generated” sticker.
Providers have technical responsibilities to make qualifying synthetic output machine-readable and detectable. Deployers have clear disclosure duties in narrower publication contexts, especially deepfakes and certain text used to inform the public on matters of public interest. Human review, editorial responsibility, standard editing and the nature of the content can all change the analysis.
That makes the first question more useful than “Was AI involved?”
Ask instead: Who is responsible, what kind of content is this, what is it being used for, and does the relevant Article 50 duty actually apply?
Once those questions are separated, AI content labelling becomes much less mysterious. The law is trying to make synthetic content and AI interactions more transparent without pretending that every use of AI creates the same risk or needs the same label.