Labelling AI content: Article 50 and exemptions

Start with meaning

The most common mistake is counting what share of the text a model wrote. The provision sets no percentage threshold, and content mixed with human-made material can still be synthetic content (European Commission guidelines, point 59). The deciding question is whether AI changed the meaning, the evidence or the apparent authenticity of the material. Under the guidelines, fixing typos and grammar and formatting are standard editing that does not trigger the marking duty in Article 50(2) (points 90 and 91), and compression and similar minor technical edits do not turn an image, audio or video into a deep fake (point 116), even if the tool passed over the whole material.

New argumentation, a cloned voice, an altered scene or a product shown with a feature it does not have is a different situation. Working categories that help start the analysis are: AI assisted, AI substantially modified, AI generated from scratch. They serve as an organising tool, while the wording of the law needs checking separately.

Gate one: who holds the obligation

The obligation rests with the company. A person using an AI tool as part of their job is not a separate deployer; the employer is. A company that merely commissions an advertisement from an agency, without deciding whether and how the agency uses AI, is not a deployer (European Commission guidelines, point 14).

The role is not fixed for good. A company is a provider when it develops an AI system or has one developed and places it on the market or puts it into service under its own name or trademark (Article 3(3)). That includes a company selling, under its own brand, a tool connected to somebody else's model through an API, even though it trains no model of its own. As a provider it is responsible for the obligations in Article 50(1) and (2) when its systems are covered by them. This distinction decides the scope of obligations, so settle it before the rest of the analysis.

Gate two: what kind of communication

Each type has its own rule and its own exceptions, so answer for each one separately.

• A system intended to interact directly with people: the provider ensures people are told they are interacting with AI at the latest at the first interaction, unless this is obvious (paragraph 1), with no editorial exception
• The system generates synthetic image, audio, video or text: the provider ensures machine-readable marking and detectability (paragraph 2), and the AI Act requires a visible disclosure from the deployer only for deep fakes and text on matters of public interest (paragraph 4)
• Emotion recognition or biometric categorisation: inform the people exposed, this overlaps with GDPR
• Deep fake: the deployer discloses that AI generated or manipulated the content, visibly or audibly and at the latest at first exposure (paragraphs 4 and 5)
• Text informing the public on a matter of public interest: disclosure, unless the editorial exception applies

Gate three: the deep fake and public interest test

For image, audio and video, the deep fake definition requires two things at once: the AI-generated or manipulated material resembles existing persons, objects, places, entities or events, and it would falsely appear to a person to be authentic or truthful (Article 3(60)). Under the European Commission's guidelines, resemblance to something that exists or could plausibly exist is enough, so a photorealistic person nobody knows also meets the first condition. The second condition is assessed in the context of publication and the audience's expectations, and photorealism alone does not settle it (points 113 and 114). An openly fantastical scene, such as a dragon or a person flying without mechanical aids, does not meet the definition. A cloned voice or face of a real person who never said the sentence meets it whenever the audience could take the recording as genuine. That person's consent to the use of their likeness does not change this, because Article 50(4) has no consent exception. Consent and a transparency label are two separate obligations.

For text there are three questions: did AI generate the text or change it beyond standard editing, is the text published to inform the public, and does it concern a matter of public interest. The European Commission's guidelines list, among others, politics and democratic processes, public administration and services, the administration of justice, fundamental rights, public security, public health, environmental protection, consumer safety, and any economic, financial, scientific or cultural development that may be a subject of public debate (points 59 and 131). Three yes answers mean disclosure is required, unless the editorial exception applies.

The editorial exception is narrower than it looks

The provision requires two things: that the material went through human review or editorial control, and that a natural or legal person holds editorial responsibility for the publication. The exception covers only text on matters of public interest. Under the European Commission's guidelines, review means a substantive check of the content by a person with relevant knowledge, including at least fact-checking. A spelling or grammar check, an automated review or a cursory sign-off does not meet it, and a substantive AI change made after sign-off removes the exception. The identity and contact details of the person or function holding editorial responsibility should be easy to find publicly, for example in the site's legal information (points 134, 135, 136 and 138).

The practice we apply during implementation goes further. We record who performed the review and when, check each substantive part separately and review the text again after any substantial edit, including a manual one. In a dispute the company has to be able to show that the review was real. The record and the part-by-part check are our caution rule, which neither Article 50(4) nor the guidelines expressly require.

Visible, audible, detectable

Machine-readable marking, meaning metadata and watermarks, falls on the system provider under Article 50(2) and must be effective, interoperable, robust and reliable as far as technically feasible. For a deep fake or text covered by Article 50(4), the deployer cannot rely on it, because the average reader does not inspect file properties. A visible or audible label in normal contact with the content is needed then (European Commission guidelines, points 117 and 132).

Where the obligation exists, the information must be clear and distinguishable, given at the latest at first exposure, and meet accessibility requirements (Article 50(5)). Under the European Commission's guidelines, information that is easy to overlook, or that the reader has to click or expand to see, does not meet that standard (points 132 and 142). The provision requires disclosing that the content has been generated or manipulated, so we avoid wording such as “this may have been generated”. A platform can crop a label while processing the material, and the deployer should ensure the audience can still see it (point 12). So check the material as an ordinary user would, on mobile and desktop, and assess it again after moving it to another channel.

Record the decision and its reasoning

From 28 October 2026 the Polish Commission for the Development and Security of Artificial Intelligence can demand documents and oral or written explanations during an inspection (Article 52 of the Polish act on artificial intelligence systems). Under the European Commission's guidelines, a company that has not signed the code of practice should be able to show by other adequate means how it complies with Article 50(4) and (5) (point 148). The record we propose covers the material identifier, the tool and model version, the AI usage category, the test outcome, the decision with its reasoning and the person who approved it, with a date. A dated screenshot is stronger evidence than a bare link, because published material can be changed or removed.

Do not copy raw prompts. A full conversation log creates a new leak surface for confidential and personal data, and proves nothing better than a decision note. Keep the record proportional to your publishing scale.