PLBook a call
Menu

AI content disclosure

Decide which content you must label, then label it so the reader actually sees it.

We prepare a procedure for a team publishing different content types. We decide when a label is needed, how it appears in each channel and who approves it. Article 50 of the AI Act applies from 2 August 2026; Poland's inspection and penalty provisions apply from 28 October 2026. Official sources are linked below.

Discuss this service ↗

30 minutes to discuss your needs and a next step. Work scope and pricing are agreed before we start.

Reply within 1 business day, scope and price proposal within 2 business days after the call, with the option to end the engagement after accepting the first stage.

Illustrative example

What the workflow can look like

  1. A catalogue of real content types
  2. A decision and reason for each
  3. Channel labels
  4. Dry run and record
01

What we establish first

We check who publishes the material and who owns each decision. We separate a tool provider's duties from the duties of the company publishing the content. Then we list formats, channels and the person who approves labels.

Article 50 of the AI Act applies from 2 August 2026. Poland's inspection and penalty provisions apply from 28 October 2026. The guide covers the legal distinctions in detail, with official sources linked below.

  • List of formats and channels
  • Provider and publisher roles
  • Decision approver
02

A classification that ends the arguments

We start from your actual content. Every material type lands in one of three categories: AI assisted the work, AI substantially modified the material, or AI created it from scratch. The deciding question is whether AI changed the meaning, the evidence or the apparent authenticity of the content.

Fixing typos, punctuation, formatting and compression leaves meaning and reality untouched. New argumentation, a cloned voice, an altered scene or a product showing a feature it does not have is a different category. The line depends on what AI did to meaning and authenticity, whatever share of the text a model wrote.

  • A catalogue of the content types you actually publish
  • A category assigned to each type
  • A list of edge cases to resolve
  • A rule for what to do when a new type appears
03

A label the reader will actually see

Machine-readable marking, meaning metadata and watermarks, falls on the system provider under Article 50(2) and does not replace information for a human. The provider of a system intended to interact directly with people must ensure disclosure of the AI interaction under paragraph 1. The deployer discloses that a deepfake, or text published to inform the public on matters of public interest, was generated or manipulated by AI (paragraph 4). The text obligation does not apply when the content underwent human review or editorial control and a natural or legal person holds editorial responsibility for its publication. Under the European Commission's guidelines, a spelling or grammar check or a cursory sign-off does not count as such review. Where the obligation exists, the information must be clear and distinguishable, given at the latest at first interaction or exposure, and meet accessibility requirements (Article 50(5)).

We draft exact wording for every channel you publish in. A chatbot informs before the first exchange. Images and posts carry a label that is visible without expanding the caption. Video gets a persistent caption, audio a spoken message. The optional EU icon can help label content, but its presence alone does not establish legal compliance. The European Commission has published Article 50 guidelines and a voluntary code of practice for marking and labelling AI-generated content. We check the messages on mobile and desktop, because platforms render differently and can crop a label.

04

What remains as evidence

From 28 October 2026 the commission can demand documents and oral or written explanations during an inspection (Article 52 of the Polish act). We build a record proportional to your scale that helps explain each decision: material identifier, tool and model version, AI usage category, test outcome, the decision with its reasoning, and who approved it and when.

We do not copy raw prompts into the record. A full conversation log creates a new leak surface for confidential data and proves nothing better than a decision note. A dated screenshot is also stronger evidence than a bare link, because published material can be changed or removed.

05

A dry run on your own publications

We work through a dozen or so materials you have actually published or plan to publish. Each gets a decision, a reason and ready wording. Disagreements between reviewers are valuable here, because they show where the rule is still unclear.

After that run the team knows who classifies, who approves and what happens with an edge case. Accountability has a named owner.

06

Scope, acceptance and pricing

Acceptance covers the content taxonomy, labels for each channel, a record template, role descriptions and a documented dry run. The materials stay with you and you can develop them without us.

Pricing depends on the number of channels, content types and how many decisions must be settled from scratch. Scope and price are agreed before work starts.

FAQ

Questions before getting started

Does content created before 2 August 2026 need labels?

Under the European Commission's non-binding guidelines (point 154), deepfakes generated or manipulated before 2 August 2026 need no retroactive label even if published later, although the Commission encourages labelling. Text generated before that date and published from 2 August 2026 to inform the public on a matter of public interest must be disclosed unless it underwent human review or editorial control and someone holds editorial responsibility for its publication.

Is a watermark on its own enough?

An invisible watermark or metadata is not enough to disclose the content to the audience. Machine-readable marking is the provider's obligation (Article 50(2)); a deployer covered by Article 50(4) must give clear and distinguishable information at the latest at first exposure. A visible label within the content, including the optional EU icon, can serve that purpose. If a platform crops the label, the deployer should ensure the audience can still see it.

How do I know this scope is current?

The dates and legal bases on this page were checked on 29 September 2026 against the consolidated text of Regulation 2024/1689 as amended by Regulation 2026/1744, European Commission documents, the Polish act on artificial intelligence systems (Journal of Laws 2026 item 1003) and the Sejm and Senate resolutions (Monitor Polski 2026 items 945 and 946). Sources are listed below.

Sources

← Company knowledge and AI governance

Let’s identify the right starting point.

Reply within 1 business day, scope and price proposal within 2 business days after the call, with the option to end the engagement after accepting the first stage.

Book a call ↗