AI content transparency for Slovak SMEs

On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act became applicable. For Slovak SMEs, this does not mean placing an “AI-generated” label on every document touched by a language model. It does mean understanding who provides the AI system, who deploys it, what kind of output reaches the public and whether people could reasonably mistake synthetic content for something authentic.

The distinction matters. A blanket label can be just as unhelpful as no disclosure: it may suggest that an entire article, recommendation or customer communication was produced without human responsibility, even when AI was used only to create an illustrative image. A precise disclosure tells the audience what is artificial, at the moment it matters, while leaving editorial accountability visible.

The European Commission updated its Article 50 guidance on 6 August and its enforcement overview on 24 August 2026. The timing makes this an operational issue, not a distant compliance project. The practical response is a small publishing control that marketing, management and external suppliers can use consistently.

Three duties that are often confused

Article 50 distributes responsibility across the value chain. SMEs should separate at least three questions:

Does the system interact directly with people? Providers must design directly interactive AI systems so that people are informed they are dealing with AI, unless this is obvious in context. A customer-facing chatbot therefore needs a clear introduction.
Can synthetic output be detected technically? Providers of systems that generate synthetic text, audio, images or video must use effective machine-readable marking where technically feasible. A customer using such a service should ask what marking survives export and ordinary processing.
Does the published material require visible disclosure? Deployers must disclose deepfakes and, in defined circumstances, AI-generated or manipulated text published to inform the public on matters of public interest.

Machine-readable provenance and a visible notice serve different audiences. Metadata may help platforms and detection tools. It does not necessarily tell a person viewing a cropped screenshot what they are seeing. Conversely, a caption visible on a webpage may disappear when the image is downloaded. A robust workflow keeps the provider's technical marking where possible and adds a clear contextual disclosure when the deployer's duty or the risk of misunderstanding calls for it.

Not every AI-assisted image is automatically a deepfake

The AI Act defines a deepfake as AI-generated or manipulated image, audio or video content resembling existing persons, objects, places, entities or events that would falsely appear authentic or truthful. A clearly stylised abstract illustration is therefore different from a photorealistic scene that could be read as a record of a real meeting, factory, accident or statement.

Context can change the assessment. A fictional office scene used as a decorative hero image may be understood as illustration, yet the same image beside a claim about a named company could imply documentary evidence. For creative, artistic, satirical, fictional or analogous works, the Act allows disclosure in an appropriate manner that does not hamper enjoyment of the work. That supports concise captions and overlays rather than oversized badges that dominate the composition.

For a small company, the safest publishing question is straightforward: could a reasonably attentive reader believe that this shows an actual person, place or event? If the answer is yes or uncertain, disclose the artificial origin clearly and close to the image. Also ensure that the wording identifies the image—not the whole article—as AI-generated.

Human editorial control changes the treatment of text

Article 50 also addresses AI-generated or manipulated text published to inform the public about matters of public interest. However, the disclosure obligation does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication.

This is not a shortcut for accepting an AI draft. Meaningful review should confirm claims against primary sources, challenge unsupported conclusions, remove invented details, assess whether the tone is misleading and establish who approves the final publication. A named internal owner or documented client approval makes responsibility operational. It also produces better content.

Routine language correction and standard editing that do not substantially alter input or meaning are treated differently from generation. Nevertheless, disclosure under the AI Act does not replace copyright, consumer-protection, data-protection, professional or contractual duties. Confidential customer information should not be placed into a public AI service merely because the eventual output can be labelled.

A practical publishing checklist for an SME

A useful control can fit on one page and remain proportionate:

Inventory the use: record the system, purpose, content type and intended channel.
Classify the role: establish whether the company is using a third-party tool, operating an interactive system or providing its own AI system.
Assess authenticity: decide whether image, audio or video could be mistaken for a real person, object, place or event.
Confirm technical marking: ask the provider what machine-readable information is embedded and whether exporting, resizing or platform upload removes it.
Choose precise visible wording: say “Illustrative image generated with AI” when that is the truth; avoid language that implies the complete publication was automated.
Place disclosure at first exposure: keep it legible, accessible and attached to the relevant content on desktop and mobile.
Review and approve: check facts, rights, personal data, misleading impressions and brand suitability; record the responsible person and date.
Preserve evidence: retain the approved output, prompt or brief, source list, licence terms and publication version for an appropriate period.

Do not let the checklist become a substitute for judgement. A product description corrected for grammar carries a different risk from a synthetic testimonial, an altered executive voice or an image implying that a customer attended an event. Higher-risk formats should receive legal review and stronger approval controls.

Turn transparency into a trust signal

Good disclosure is not an apology for using technology. It allows a business to benefit from efficient tools while being honest about what customers see. Precise language can even strengthen the distinction between a synthetic illustration and the human expertise behind the article, offer or recommendation.

The process should also be tested in the real channel. Check whether a social platform crops the notice, whether a mobile layout covers it, whether assistive technology can access the information and whether downloaded assets retain useful provenance. Procurement questions, content templates and a final pre-publication check are more reliable than expecting every employee to interpret the regulation from scratch.

For more on building responsible digital workflows, see our Technology services.

Sources and further reading

Legal and factual statements were checked against the following official sources on 30 August 2026. The checklist, examples, practical recommendations and wording are original editorial work by Merkle s. r. o. This article provides general information and does not replace legal advice for a specific case.