Do you have to label a LinkedIn post made with AI?

A typical LinkedIn post prepared with a model’s help and checked by a person usually needs no label. Check whether you are publishing a deep fake or text informing the public on a matter of public interest. Article 50 of the AI Act sets out those cases.

The provision does not count percentages

Article 50 of Regulation 2024/1689 has applied since 2 August 2026. In Poland, the act on artificial intelligence systems entered into force on 11 August 2026, and the rules on inspections, proceedings and penalties apply from 28 October 2026 (Journal of Laws 2026 item 1003, Article 127).

The deciding question concerns meaning: did AI change the sense, the evidence or the apparent authenticity of the content. How much of the text a model wrote sits on a completely different axis. You can run an entire text through a tool without triggering the obligation, and you can change one sentence and trigger it fully.

Fixing typos, punctuation, formatting and compression leaves meaning untouched. So does research where a model suggests a direction and a person verifies the sources and writes. New argumentation nobody checked, a cloned voice or a product shown with a feature it does not have is a different category.

For an ordinary B2B post, usually not

A typical LinkedIn post about how your company solved a customer problem, written by a person with a model’s help and checked by that person, is neither a deep fake nor text informing the public on a matter of public interest. It therefore passes through none of the Article 50 gates.

The situation changes in four places, and those deserve more attention than the post itself:

Graphics with a synthetic person or an altered product. Context decides here, whatever generator was used. The deep fake definition requires resemblance to an existing person, object, place or event plus a false appearance of authenticity. An avatar resembling a real person, including a photorealistic AI-invented person who could plausibly exist, or a real product shown with a feature it does not have, usually meets those conditions if the audience could take the material as genuine (European Commission guidelines, points 113 and 114). A wholly fantastical figure with no reference to reality usually does not. Invented product features carry separate consumer protection risk.

Video with a cloned voice or face. A recording in which your CEO says a sentence they never said is a deep fake whenever the audience could take it as genuine. The European Commission’s guidelines give a similar example: a video with a realistic synthetic avatar of a company CEO congratulating employees on the year’s results (point 114 and the examples after point 116). That person’s consent to the use of their likeness changes nothing here, because Article 50(4) has no consent exception, and consent and a transparency label are two separate obligations.

A chat on your website. Article 50(1) requires systems to be designed so a person knows they are interacting with AI, at the first interaction at the latest. The obligation is addressed to the system provider, and the provision carries an exception: no notice is needed where this is obvious to a reasonably well-informed observer. In practice, if you deploy somebody else’s chat under your own brand, check your contract for who owns that message.

Text on matters of public interest, for example health, safety, the environment, politics, regulation or significant economic events. If the publication is meant to inform the public and AI substantially changed the content, the obligation appears unless the editorial exception applies.

The editorial exception is narrower than it sounds

This is where mistakes happen most easily, because the name suggests something far broader. The exception covers only text on matters of public interest, and requires two things at once: that the material went through human review or editorial control, and that a natural or legal person holds editorial responsibility for the publication.

Under the European Commission’s guidelines, neither a review by a second model nor a quick glance before clicking publish meets that condition, and a substantive AI change made after the review removes the exception (points 135 and 136). Beyond that we apply caution: we check each substantive part separately and we restart the review after any substantial manual edit too. In a dispute, the company has to be able to show the review was real.

When a visible label is needed

The second common confusion. Machine-readable marking, meaning metadata and watermarks, is a separate obligation on the system provider and does not replace information for a human. Nobody opens file properties before reading a post.

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 material may have been generated”. Platforms can crop a label while processing an image, and the deployer should ensure the audience can still see it (point 12). So look at your own material the way an ordinary reader sees it, on mobile and desktop, and assess it again after moving it to another channel.

For the full decision sequence by content type, duty and exception, read the guide to labelling AI content under Article 50.

What to do this week

This does not need an implementation project. List the content types you actually publish, assign each one a decision with its reasoning, and settle who approves it. For materials that need a label, record the identifier, the tool used, the category, the decision and the approval date.

If you publish across several formats or channels, see our AI content labelling implementation for a shared classification procedure, messages and team record.

One thing should stay out of that record: raw prompts. A full conversation log creates a new leak surface for confidential data and proves nothing better than a decision note.