Not all content made with artificial intelligence needs to receive the same label, and in some cases there is no obligation to display a publicly visible warning.
Since August 2, 2026, article 50 of the AI Act has started to require different forms of transparency for systems that talk directly to people, tools that generate synthetic content, deepfakes and certain texts on matters of public interest. The obligation may fall on whoever provides the system, on whoever uses it professionally, or on both.
The most important distinction is this: a machine-readable technical mark is not the same thing as a reader-visible label. There is also no general rule determining that every text, image or video that has received some help from AI displays the phrase “generated by artificial intelligence”.
What actually changed on August 2, 2026
The European Artificial Intelligence Regulation, known as the AI Act, was approved in 2024 and had its application divided into stages. Article 50, which deals with transparency obligations for certain AI systems, became applicable on August 2, 2026.
The guidelines used to interpret these obligations, however, were not published as of August 2. The European Commission published them on July 20, 2026, before the rules came into effect. Therefore, the event of August 2nd was the practical validity of article 50, and not the original publication of the guidelines.
From that date onwards, the rules reach four main situations:
- direct interaction between a person and an AI system;
- generation or manipulation of audio, image, video or synthetic text;
- use of emotion recognition or biometric categorization;
- exposure to deepfakes or certain texts of public interest produced or manipulated by AI.
There is a limited transition: content generating systems that had already been placed on the market or in service before 2 August 2026 have until 2 December 2026 to specifically meet the marking and detection obligation under Article 50(2). This extension did not postpone all other obligations of article 50.
Content generated and made available before August 2, 2026 also does not need to be retroactively labeled, although the Commission encourages voluntary transparency when possible.
Technical marking and visible warning are different obligations
The AI Act separates the responsibilities of providers and deployers, called providers and deployers in the English version.
The provider is, in general terms, the person who develops an AI system or has it developed and places it on the market or into service under its own name or brand. The implementer is the company, agency or professional who uses this system under their authority. Personal and non-professional use falls outside this definition of implanter.
Machine-readable marking
Providers of systems that generate audio, image, video, or synthetic text must structure the outputs so that they are marked in a machine-readable format and discoverable as generated or manipulated by AI.
The solution must be effective, interoperable, robust and reliable to the extent that this is technically feasible. The rule allows limitations related to the type of content, the state of the art and the cost of implementation.
This obligation does not apply when the AI performs only an ordinary auxiliary editing function or does not substantially change the input data or its meaning. The Commission also mentions narrow exceptions for some machine-to-machine-only output, source codes, short sequences, and certain closed industrial environments.
People-perceivable warning
In other situations, it is not enough to insert a technical tag in the file. The notice must be clear, distinguishable and presented, at the latest, upon the person's first contact or first exposure to the system or content. It must also comply with applicable accessibility requirements.
This occurs, for example, in interactions with chatbots and the dissemination of deepfakes. For a deepfake, the Commission expressly states that the deployer cannot rely solely on a technical marking embedded by the provider: the information needs to be perceptible by a person, such as a visual or audible warning.
Practical matrix: when is it necessary to inform or label?
| System or content type | Who is responsible for transparency | What the rule requires | Main limits or exceptions | What the rule does not allow to conclude |
|---|---|---|---|---|
| Chatbot or AI agent | As a rule, the system provider | Inform, from the beginning of the first interaction, that the person is talking to an AI. The obligation can be waived when it is really obvious in the context. | Background systems, communication only between machines or situations without direct contact with a person do not fall under this specific obligation. The exception of the “obvious” character must be interpreted restrictively. | The warning does not mean that the system is high risk, that all the answers are wrong or that it complies with all other standards. |
| Interactive Avatar | Provider; in some situations, also the deployer | If the avatar maintains a direct, two-way interaction with a person, it must be clear that it is an AI system. If it also generates synthetic content, additional technical marking obligations may exist. | A character only displayed in a video, without direct interaction, does not automatically fall under the chatbots rule. If you imitate a real person, object, place or event in a seemingly authentic way, you must evaluate the deepfake rule. | Not all avatars are deepfakes. Human appearance alone does not determine that the content is attempting to impersonate a real person or event. |
| Entirely synthetic image | Generator system provider | The output must, as a rule, contain machine-readable markup that allows AI generation or manipulation to be detected. | Marking may not be required when the tool only performs ordinary editing or does not substantially alter the input material. Systems prior to August 2, 2026 are transitioned until December 2, 2026 for this specific obligation. | The rule does not require that every synthetic image has a visible stamp. An image created from scratch also does not become a deepfake just because it was made with AI. |
| Deepfake in image, audio or video | Implementer who uses or disseminates the content; the provider may have a parallel technical marking obligation | Inform clearly and perceptibly that the content was artificially generated or manipulated, at most on the first exposure. | In an obviously artistic, creative, satirical or fictional work, the notice may be presented in a way that does not harm the exhibition or enjoyment of the work. There is an exception for legally authorized uses in certain criminal investigation activities. | The label indicates generation or manipulation and risk of inauthentic appearance. It is not, in and of itself, a ruling that every message in the content is false, illegal or defamatory. |
| Public interest text without substantive human review | Implementer that publishes the text | Clearly state that the text was generated or manipulated by AI when it is published to inform the public about a matter of public interest. | The concept includes topics such as politics, public services, Justice, fundamental rights, security, public health, environment, consumer protection and economic, financial, scientific or cultural developments relevant to public debate. | The obligation does not automatically apply to all commercial, fictional, internal or private text. The label also does not replace investigation, attribution of sources or correction of errors. |
| Text reviewed by a person with editorial responsibility | Implementor or entity responsible for publication | Article 50(4) dispenses with the visible label when there has been human review or substantive editorial control and a natural or legal person has assumed editorial responsibility for the publication. | Spelling correction, grammatical adjustment or merely formal verification do not count as substantive revision. The responsible person must be able to examine, modify, reject and verify the content and its sources. | Dispensing with the label does not certify that the text is correct. It also does not eliminate any technical marking inserted by the provider or remove copyright, data protection, consumer standards or journalistic duties. |
What counts as a deepfake under the AI Act
The regulation defines deepfake as image, audio or video content generated or manipulated by AI that resembles existing people, objects, places, entities or events and that may falsely appear to be authentic or true.
This is more specific than “any image made with AI”. An admittedly fantastic illustration, for example, can be synthetic content without necessarily qualifying as a deepfake. An audio that convincingly reproduces the voice of a real person, making it appear that they uttered a certain statement, is more likely to be included in the definition.
The evaluation considers the degree of similarity, the message transmitted, the predictable context of dissemination and the public's expectations. Special effects or artificial scenery in a work from which no one expects documentary authenticity may receive different treatment from a recording presented as an actual record.
When a text of public interest can be left unlabeled
The exception for human review is one of the most relevant parts for newsrooms, websites, agencies, consultancies and communication teams.
According to the European Commission, review must involve deliberate examination of the substance of the content by one or more persons with appropriate knowledge and professional judgment. Editorial control presupposes real authority to approve, change or reject the text, check information and evaluate the reliability of sources.
So it's not enough:
- run a spell checker;
- ask the AI to “review its own text”;
- read superficially before publishing;
- change title, punctuation or some words;
- assume the AI platform has verified the facts.
The simple act of writing a prompt also does not, by itself, amount to the substantive review required for the exception. The organization needs to demonstrate that there was actual editorial intervention and that someone took ultimate responsibility for the publication.
Even in this case, the exemption applies to the visible label of the text of public interest provided for in article 50(4). It does not necessarily erase the technical marking applied by the system provider pursuant to article 50(2).
Is a European seal mandatory?
There is no single mandatory visual badge for all content made with AI.
The European Union has provided icons that can be used to identify content entirely generated or partially modified by AI. The use of these icons is optional. Legal transparency obligations, where applicable, are not optional but may be met by other appropriate forms of notice.
Using the icon also does not alone prove full compliance with the AI Act. The person responsible still needs to evaluate the content, context, position of the notice, accessibility and clarity for the public.
The European Code of Practice offers voluntary measures to demonstrate compliance. Adhering to the code is not mandatory, and companies that do not sign it remain free to demonstrate compliance with article 50 by suitable alternative means.
Verification flow: do you need to inform or label?
Before publishing or putting a system into operation, the organization can follow this sequence:
- Is there direct, two-way interaction with a person?
If so, state that it is AI from the beginning, unless this condition is unequivocally obvious. - Does the system generate or manipulate audio, image, video, or synthetic text?
Check whether the provider has implemented machine-readable technical markup and whether any auxiliary editing exceptions apply. - Does the visual or audio content appear to authentically represent a real person, object, place or event?
Evaluate whether it qualifies as a deepfake. If so, the deployer must provide a notice that is noticeable to the public. - Is this text published to inform the public about a matter of public interest?
Record whether it was generated or manipulated by AI. - Was there substantive human review or editorial control?
Document who reviewed, what facts and sources were checked, what changes were made, and who took final responsibility. - Does the notice appear on first contact or first exposure?
Confirm clarity, readability, accessibility and permanence of the notice when content is shared or reproduced.
What an AI label doesn't prove
The presence of a warning does not automatically mean that the content is false. A synthetic image can represent a correctly identified simulation; a labeled text may contain verified information; and a deepfake can be part of an artistic work clearly presented as fiction.
Similarly, the absence of a visible label does not prove that no AI was used. The involvement of technology may have been limited to ancillary editing, there may be only machine-readable technical markup, or the text may have received substantive human review and editorial responsibility.
Article 50 creates duties of transparency regarding the interaction or artificial origin of content. It does not function as a certification of veracity, quality, safety, authorship, legality or respect for third party rights.
When the AI Act can reach a Brazilian company
The fact that the legislation is European does not mean that only companies based in the European Union need to evaluate it.
The AI Act reaches providers who place AI systems or models on the European market, even when established in another country. It also provides for application to providers and implementers located outside the Union when the output produced by the system is used within the block.
A Brazilian company may need specific analysis, for example, when:
- offers a chatbot, agent or content generator in the European market;
- provides a solution under its own brand to customers located in the European Union;
- professionally operates a system whose results will be published, displayed or used in the block;
- produces synthetic content for campaigns, services or publications aimed at the European market;
- acts simultaneously as a provider of technology and as responsible for its use.
This does not mean that any Brazilian website accessible abroad is automatically subject to all obligations. The assessment depends on the role of the organization, the system's offering, the market involved, the use of the output and the concrete circumstances of the operation.
Companies with users, customers, operations or content distribution in the European Union should map their flows and seek expert legal review before concluding that they are within or outside the territorial reach of the regulation.
Inspection and possible fines
The application of Article 50 will be conducted mainly by the national market surveillance authorities of European countries. The AI Office has competences in specific situations, especially for certain systems based on general-purpose AI models, while the European Data Protection Supervisor acts over Union institutions and bodies.
Non-compliance with transparency obligations may be subject to administrative fines of up to 15 million euros or, for companies, up to 3% of the previous year's global annual revenue, depending on the criteria and limits of the regulation. The legislation provides for proportionality in the application of penalties, including considerations for smaller companies.
The minimum editorial checklist
For each professional use of AI, record:
- which system was used;
- who is the provider and who is the deployer;
- if there was direct interaction with people;
- what type of content was generated or manipulated;
- if there is a technical marking;
- whether the material can be confused with an authentic record;
- if the text deals with a matter of public interest;
- who performed the substantive review;
- who assumed editorial responsibility;
- which warning was displayed and at what time;
- what evidence of compliance has been preserved.
Save this verification box before publishing content produced with AI. To complement the analysis, also consult the related guide on how to check allegations involving artificial intelligence agents.
