The European Commission has published new guidelines explaining how companies should comply with the EU AI Act transparency rules in Article 50. The rules are designed to make it easier for people to recognize when they are interacting with artificial intelligence or viewing content that was generated or manipulated by AI.
Most of these obligations take effect on August 2, 2026. They cover interactive systems such as chatbots and AI agents, machine-readable marking of synthetic media, visible disclosures for deepfakes, and labels for certain AI-generated texts on matters of public interest.
Although this is European legislation, its reach is not limited to companies based in Europe. According to the European Commission’s official Article 50 guidance, providers outside the European Union may also be covered when the output of their AI systems is used in the EU.
Here’s what the new EU AI Act transparency rules mean in practice.
Why the EU introduced new AI transparency rules
Generative AI can now produce convincing text, images, audio, and video in seconds. Interactive systems are also becoming more human-like, making it less obvious whether a person is communicating with another human or with software.
That progress creates useful new tools, but it also increases the risks of impersonation, fraud, misinformation, manipulation, and consumer deception. The purpose of Article 50 is not to prohibit AI-generated content. Instead, it aims to give people enough information to judge what they are seeing, hearing, or interacting with.
The European Commission’s transparency guidelines clarify the scope of the law and provide examples of systems and content that are either covered or excluded.
Providers and deployers have different responsibilities
The AI Act separates the organizations involved into two main groups: providers and deployers.
A provider is generally the company that develops an AI system or places it on the market under its own name or trademark. Providers are responsible for building the required transparency features into the system.
A deployer is a person or organization using an AI system under its authority in a professional context. This can include businesses, public agencies, publishers, advertising firms, and freelancers. Employees working under a company’s authority are not normally treated as separate deployers; responsibility remains with the organization controlling the use of the system.
Purely personal and nonprofessional use is generally outside this part of the AI Act. However, activity connected to a business, trade, freelance occupation, or regular economic benefit may qualify as professional use.
Chatbots and AI agents must identify themselves
Providers of systems that interact directly with people must ensure users are informed that they are communicating with AI. Examples include:
- Customer-service chatbots
- AI agents
- Virtual assistants
- Conversational avatars
The notice must appear in a clear and distinguishable way from the beginning of the first interaction. It should also meet accessibility requirements.
There is an exception when it is already obvious to a reasonably well-informed and observant person that the interaction involves AI. However, the Commission says this exception should be interpreted narrowly. A company should not assume that every user will automatically recognize a realistic voice agent, avatar, or conversational interface as artificial.
Systems operating entirely in the background, machine-to-machine communications, and tools without direct contact with a person generally fall outside this specific disclosure requirement.
AI-generated content needs machine-readable marking
Providers of generative AI systems must make synthetic or manipulated text, images, audio, and video detectable as AI-generated. This requires an effective, reliable, robust, and interoperable machine-readable mark, as far as technically feasible.
This mark may function like a digital watermark or embedded metadata, but it does not necessarily have to be visible to the human eye. Its purpose is to allow platforms, software, and verification tools to detect the artificial origin of the content.
The requirement focuses on the design of the AI system. In other words, providers must equip their products with a technical method for marking and detecting generated or manipulated outputs.
Deepfakes require a visible or audible disclosure
Deployers have an additional responsibility when they publish deepfakes. Under the AI Act, a deepfake is AI-generated or manipulated image, audio, or video content that resembles an existing or plausibly existing person, object, place, entity, or event and could falsely appear authentic or truthful.
In these cases, the disclosure must be understandable and perceivable by a person no later than their first exposure to the content. Depending on the format, this could be a visible label, an on-screen notice, or an audible warning.
A hidden machine-readable mark is not enough by itself. The technical marking added by the AI provider and the public-facing disclosure added by the deployer serve different purposes.
The rules are more flexible for content that is clearly artistic, fictional, satirical, creative, or similar. Disclosure is still required when the material qualifies as a deepfake, but it can be presented in a way that does not interfere with the experience of the work.
When AI-generated text must be labeled
The rules do not require every piece of AI-assisted writing to carry a public label. The obligation applies to AI-generated or AI-manipulated text when all three of the following conditions are met:
- The text is published.
- It is intended to inform the public.
- It concerns a matter of public interest.
The Commission lists areas such as politics, democratic processes, public administration, justice, law enforcement, fundamental rights, public safety, public health, environmental protection, consumer safety, economics, finance, science, and culture.
There is an important exemption. A publication does not need this label if it has undergone meaningful human review or editorial control and a natural or legal person accepts editorial responsibility for it.
The review must examine the substance of the content. Fact-checking, validating sources, correcting factual claims, and having an editor who can approve, revise, or reject the text can qualify. A basic spelling or grammar check does not.
This distinction matters to newsrooms, company blogs, public-sector communications teams, and other organizations that use AI during content production. Simply opening an AI-generated draft and fixing a few typos is unlikely to meet the Commission’s standard for substantive human review.
Emotion recognition and biometric categorization
Deployers must also inform people when they are exposed to emotion-recognition or biometric-categorization systems. The obligation applies whether the analysis occurs in real time or after the data has been collected.
This rule is separate from the obligations covering generative content. It addresses situations in which AI analyzes personal characteristics or attempts to infer emotional states without the affected person necessarily realizing that the system is operating.
What falls outside the marking requirement?
The guidelines identify several outputs and uses that may fall outside the machine-readable marking obligation, including:
- Short sequences of numbers, letters, or symbols
- Source code
- Content created solely for automated machine-to-machine communication
- Certain outputs used only inside closed-loop industrial or product-development environments
- Standard editing that does not substantially alter the original data or its meaning
Standard editing can include limited assistive tasks such as spelling and grammar corrections. The exact assessment depends on what the system changes. Once AI materially changes the substance, meaning, or presentation of the input, the exception may no longer apply.
The guidelines also describe a narrow exemption for some business-to-business and industrial uses, provided that all relevant conditions are met. Companies should therefore evaluate the actual workflow instead of assuming that any internal or B2B use is automatically excluded.
Key dates and the limited grace period
Article 50 applies beginning August 2, 2026. From that date, providers and deployers must comply with the relevant transparency obligations.
There is a limited grace period for generative AI systems placed on the market before August 2, 2026. For those systems only, the provider’s machine-readable marking and detection obligation begins on December 2, 2026.
The delay does not suspend all of Article 50. It applies only to the technical marking requirement for qualifying systems already on the market.
Content generated before August 2, 2026 does not have to be labeled retroactively. The Commission nevertheless encourages voluntary disclosure when possible.
The Code of Practice is voluntary, but the law is not
The EU has also published a Code of Practice on Transparency of AI-Generated Content. Signing the code is voluntary, but complying with Article 50 is mandatory.
Providers and deployers that sign the code can rely on its measures as a recognized way to demonstrate compliance with the marking and labeling obligations. Organizations that choose another approach must be able to show that their alternative measures are equally adequate.
The EU has also released optional icons that deployers can use to label AI-generated content. These can help create a more consistent visual language, but using an icon does not replace the need to evaluate whether a particular system, output, or publication falls within the law.
Penalties for violating the transparency obligations
Enforcement will primarily be handled by national market-surveillance authorities. The AI Office has authority in certain cases involving systems based on general-purpose AI models, while the European Data Protection Supervisor is responsible for AI systems used by EU institutions and bodies.
According to the Commission’s quick facts on AI transparency, fines can reach €15 million or 3% of a company’s total worldwide annual revenue for the preceding financial year. Proportionality rules can be considered for small and medium-sized businesses and small mid-cap companies.
What businesses should do now
Organizations using generative or interactive AI should start by mapping where these systems appear in their products and workflows. A practical review should answer several questions:
- Do customers interact directly with a chatbot, voice agent, or AI avatar?
- Can the system create or modify public-facing text, images, audio, or video?
- Does the organization publish deepfakes or realistic synthetic media?
- Is AI used to produce text about matters of public interest?
- Who performs substantive human review and accepts editorial responsibility?
- Are emotion-recognition or biometric-categorization tools in use?
- Can the provider supply reliable machine-readable marking and detection features?
Companies should also document their review procedures, assign ownership for disclosures, test whether labels remain visible across platforms, and verify what their AI vendors provide. For publishers, maintaining a genuine editorial process will be especially important when AI assists with public-interest content.
US companies should not ignore these obligations simply because they are headquartered outside Europe. If an AI system’s output is used in the EU, the business may still fall within the scope of the AI Act.
A new baseline for AI transparency
The EU AI Act transparency rules establish a basic principle: people should be able to tell when AI is participating in an interaction or shaping content that could influence what they believe.
For providers, that means building disclosure and machine-readable marking into AI systems. For professional deployers, it means clearly notifying people about deepfakes, certain public-interest texts, emotion recognition, and biometric categorization.
The rules will not eliminate deceptive synthetic media, and technical marks can still face practical limitations. However, Article 50 creates a common legal baseline for greater accountability. With the August 2026 deadline approaching, companies that rely on AI should review their systems, vendors, content workflows, and editorial controls now.
This article provides general information and does not constitute legal advice.
