AI regulation · Updated 6 October 2026
Information, not legal advice. Check your own situation with counsel who knows EU media and AI law.
The short answer. You need a visible disclosure if four things are all true: you use an AI system professionally, the text it generates or manipulates is published, the text aims to inform the public on a matter of public interest, and no human review exemption applies. That exemption only works when a competent person substantively checks the text and a named person carries editorial responsibility. A spell-check does not count. Most affiliate reviews, novels and internal documents sit outside the rule; AI-written news summaries, explainers on health, finance, law or politics usually sit inside it. In this guide
- What Article 50(4) says
- The four gates
- Where “AI-assisted” lands
- The human review exemption
- How to label, and what fails
- Dates, enforcement, fines
- A working checklist
- What is still unsettled
- Sources
What Article 50(4) says
Article 50 of the EU AI Act (Regulation 2024/1689) collects the Act’s transparency duties. Paragraph 4 puts two duties on deployers, meaning whoever uses an AI system under their own authority for something other than personal, non-professional activity. The first covers deepfakes: realistic AI-made image, audio or video that would falsely appear authentic. The second, which matters for publishers, covers text: if an AI system generates or manipulates text that is published to inform the public on matters of public interest, the deployer must disclose that the text is artificially generated or manipulated.
The statutory wording is short. Almost everything useful for a site owner comes from the Commission’s guidelines on Article 50, its Q&A page and the Code of Practice on transparency of AI-generated content. The guidelines are interpretive and not binding law; courts and national regulators have the last word. They are still the best map of how enforcers are likely to read the text.
The four gates
The Commission describes the text rule as a set of cumulative conditions. Fail any one and Article 50(4) does not require a label for that piece.
- You are a deployer. A person who uses AI casually on a private blog with no income is outside the Act. A business, freelancer, or anyone earning from the activity on a regular basis counts as professional. Staff acting under a company’s instructions are not separate deployers; the company is.
- The text is published. Per the guidelines, that means reachable by an open-ended, fairly large group of unrelated readers, paid or free. Private emails, internal documents and small closed groups are not published.
- It is meant to inform. The text must aim to convey facts, knowledge or opinions. Entertainment fiction and poetry do not.
- Public interest. The Commission’s examples include politics and democratic processes, public services, justice, fundamental rights, public security, public health, environment, consumer safety, and economic, financial, scientific or cultural developments that could be a subject of public debate.
Then comes the exemption: even a text that passes all four gates needs no label if it went through human review or editorial control and someone holds editorial responsibility.
Where “AI-assisted” lands
The Act says “generates or manipulates,” which is wider than “AI wrote it from scratch.” The guidelines give examples mainly for the provider-side marking rule in Article 50(2), where standard editing is excluded: grammar and spelling correction, minor polishing and formatting stay out, while AI summaries and rewriting that changes style, structure or meaning stay in. Applying that line to Article 50(4) is a reasonable reading, but it is an inference, not an official ruling. The table below is my reading, not the Commission’s.
| What you publish | Likely label needed? | Why |
|---|---|---|
| Human-written article, AI fixes typos and grammar | No | Looks like standard editing; meaning and style unchanged |
| AI-generated news summary on your site | Yes, unless exempt | The guidelines list AI news summaries as in scope |
| AI draft on public health or finance, lightly skimmed and posted | Yes | Skimming is not substantive review |
| AI draft, then fact-checked and approved by a competent editor who is publicly named | No | Both exemption conditions met |
| Human-reviewed article, then AI rewrites sections afterwards | Yes, for the changed text | Substantive AI intervention after review can void the exemption |
| AI-written novel, poem, or entertainment piece | Not under 50(4) text rule | Not published to inform |
| AI product blurb with claims on health, safety or sustainability | Possibly | Advertising is out of scope unless it informs on public-interest matters |
| Chatbot answer shown only to the one user who asked | No, but other rules may apply | Not published to the public |
| Internal memo or small private group message | No | Not published |
The human review exemption, in detail
This is where most sites will win or lose. The Commission sets two conditions, and both must be true.
1. Real human review or editorial control
Human review means a person (or several) with relevant knowledge and professional judgment deliberately examines the substance of the content. Editorial control means a responsible editorial entity, such as an editor-in-chief, has authority in practice to approve, change or reject the text on substantive grounds, including fact-checking and source reliability. The guidelines say superficial, formal or procedural checks, spell-checking and grammar correction among them, do not qualify. Automated review does not qualify either.
2. Editorial responsibility
Someone must hold ultimate legal responsibility for publishing the content. Under the guidelines, the identity and contact details of that natural person need to be publicly available. The Code of Practice asks non-media deployers to keep a policy naming the person or function responsible and the organisational measures behind it, and to publish contact details. Media service providers can rely on procedures they already have.
Practical consequence: an anonymous byline, a persona with no accountable owner, or a “reviewed by our team” line with no named responsible person is a weak foundation. If the exemption is your plan, be ready to show what the review involved, who did it, and when.
How to label, and what fails
Article 50(5) sets the presentation standard for every Article 50 disclosure: clear, distinguishable, given no later than first exposure, and accessible.
- Placement. The guidelines expect the disclosure at the start of the publication, not buried at the end. The Code of Practice suggests the top of the text, next to the headline, or in the colophon, in a consistent way.
- No effort by the reader. The reader must not need special tools or deliberate actions to see it. A tooltip that appears only on hover is risky.
- What fails. A line only in the terms and conditions, only in end credits, or hidden behind menus does not meet the standard.
- Icons. The AI Office published three optional EU icons: “AI generated” for fully AI-made content, “AI modified” for human content partly changed by AI, and a basic AI icon. Using them is voluntary and does not by itself prove compliance. Testing reported by the Commission found that icon variants with a text label were understood better.
- Accessibility. The guidelines tie accessibility to EU law, including the European Accessibility Act and the Web Accessibility Directive. The Code of Practice points to WCAG 2.1 Level AA as a reference standard.
A minimal, screen-reader-friendly pattern, placed above the article body:
<p class="ai-disclosure" role="note">
<strong>AI-generated text.</strong> This article was produced
with an AI system and published without human editorial review.
</p>For text that was drafted by AI and then reviewed, a different notice is accurate: say that AI was used, name the responsible editor, and link to your editorial policy. Wording must match what actually happened. A label that overstates or understates the process creates its own risk.
Dates, enforcement and fines
- 2 August 2026: Article 50 applies. The only grace period the Commission describes covers machine-readable marking by providers of systems already on the market before that date, which have until 2 December 2026. It does not cover deployer labelling under Article 50(4).
- Older content: text generated before 2 August 2026 does not have to be labelled retroactively, though the Commission encourages it where possible.
- Who enforces: mainly national market surveillance authorities. The AI Office has a narrow role, for example for systems built on general-purpose models by the same entity, or those integrated into very large platforms and search engines.
- Fines: up to 15 million euros or 3% of worldwide annual turnover, whichever the Act applies, with proportionality for SMEs and small mid-caps.
- Code of Practice: finalised 10 June 2026 and assessed as adequate by the Commission and the AI Board. Around 190 organisations had signed by the end of July. Signing is voluntary, and signatories can rely on it to show compliance; others must show their own measures are adequate and may face more information requests.
A working checklist
- Inventory. List every published text where an AI system generated or rewrote more than grammar.
- Classify. Run each through the four gates. Record the reasoning, since a regulator may ask.
- Choose a path per category: label it, or route it through documented substantive review.
- Name the responsible person and publish contact details on an editorial policy page.
- Define “review” in writing: who qualifies, what they check (claims, sources, numbers), and what gets logged.
- Freeze after review. Re-trigger review if AI changes the text afterwards.
- Build the label as a template component at the top of the page, with accessible markup.
- Revisit quarterly. Guidance and national practice are still forming.
What is still unsettled
- Guidelines are not law. A national authority or court can read the Act differently.
- “Public interest” has soft edges. The Commission’s list is illustrative. Product reviews, lifestyle advice and SEO explainers fall in a grey zone, especially when they make health, safety or financial claims.
- Where the standard-editing line sits for text under Article 50(4) specifically has not been spelled out the way it has for providers.
- Territorial reach for non-EU publishers. The Commission’s Q&A states that non-EU providers are covered when their output is used in the EU. Confirm how this applies to a non-EU deployer with EU readers before assuming either way.
- Enforcement practice is only weeks old; no public enforcement pattern exists yet.
Sources
- European Commission, Transparency obligations under Article 50 of the AI Act (Q&A), updated 24 July 2026.
- European Commission, Guidelines on transparency obligations for providers and deployers of certain AI systems, updated 29 July 2026.
- European Commission, Code of Practice on Transparency of AI-generated Content, updated 31 July 2026.
- Orrick, EU AI Act: Transparency Obligations for AI-Generated Content (Article 50), 14 August 2026 (secondary analysis mapping the Act, guidelines and Code).
- Regulation (EU) 2024/1689, Official Journal text, Article 50.