Home / Insights / EU AI Act

EU AI Act: What the New AI Content Labelling Rules Mean for Your Business

By Paul O'Dwyer, Founder, Irish AI Consultancy · Published 3 August 2026 · 7 minute read

If you have used AI to help write a blog post, generate marketing images, or build a chatbot for your website, this is worth seven minutes of your time. Part of the EU AI Act, the legislation that has been building toward full enforcement for two years, became binding law on 2 August 2026. This guide explains what actually changed, who it affects, and what to check in your own business.

When did the EU AI content labelling rules come into force?

The transparency obligations sit in Article 50 of the EU AI Act, and they became enforceable on 2 August 2026. Until that date, disclosure of AI-generated content was largely a voluntary, platform-by-platform choice, the kind of thing Instagram or TikTok built their own synthetic media badges around. From 2 August, it becomes a legal duty enforced by national market surveillance authorities in every EU member state.

A separate, related duty lands later: providers of generative AI tools must build in machine-readable markers, such as watermarks or metadata, so their systems' output can be detected as artificial. For systems already on the market before 2 August 2026, that provider duty applies from 2 December 2026. Systems launched after 2 August 2026 are expected to comply from day one.

Who does this apply to?

Two groups, and most businesses using AI fall into one or both. Providers are the companies that build and supply AI systems. Deployers are the businesses that use those systems in their own products, services, or publishing, which covers the vast majority of Irish SMEs using AI tools rather than building them.

Geography offers no shelter. A provider or deployer established anywhere in the world is covered whenever its AI system's output is used inside the EU, according to the European Commission's own guidance. An Irish business, or a business anywhere else, publishing AI-assisted content that reaches EU readers is within scope.

What actually has to be labelled?

Does my business need to label AI-written blog posts?

This is the question most relevant to Irish SMEs, and the answer turns on one thing: genuine human editorial review. The Commission's draft guidelines on Article 50 set out an exemption for text specifically. If a clearly identifiable person reviews AI-generated text before publication, genuinely checking it for accuracy, plausibility and sources, with a real possibility of amending or rejecting it, and that person or the business takes editorial responsibility for the result, the text does not need a label.

The guidelines are explicit about what does not qualify: a quick spelling check is not sufficient review. Content simply "skimmed" before publishing is not exempt. What the exemption is really asking for is what a responsible publisher already does, someone accountable reads it properly and stands behind it before it goes out.

This exemption does not extend to deepfakes. A synthetic image or video that would appear authentic to an ordinary viewer needs a label regardless of how carefully it was reviewed.

What are the penalties?

Breaches of the transparency obligations can carry fines of up to 15 million euro or 3 percent of a company's global annual turnover, whichever figure is higher. Enforcement sits with each member state's market surveillance authority, and the EU AI Office has said it will run dedicated task forces from September 2026 to help providers and deployers align on practical implementation.

What should Irish businesses actually do now?

Start with a short audit of where AI already touches customer-facing content: blog posts, marketing copy, product images, any chatbot or virtual assistant on your website. For text, the practical fix is usually process rather than technology: make sure a named person genuinely reviews AI-assisted content before it publishes, and can show they did. For any chatbot or AI assistant customers interact with, add a simple, visible disclosure that they are talking to AI.

If you are building or commissioning a generative AI product rather than just using one, the provider-side marking duty is worth planning for now even though it is not enforceable until December, since retrofitting watermarking or metadata into a shipped product is harder than designing it in from the start.

The technical standards for watermarking and machine-readable marking are still being finalised through a voluntary Code of Practice, published in final form on 10 June 2026. It is worth tracking as implementation detail firms up, but it does not change the core August obligations around labelling and disclosure.