/

Updated 18 November 2026

Do You Need to Disclose AI-Generated Content? The Article 50 Test for Websites

AI Disclosure: This content includes AI-generated text.

Short answer: sometimes, and far less often than most “AI disclosure” advice suggests. Article 50 of the EU AI Act has applied since 2 August 2026. For a typical website it creates labelling duties in three situations: text on matters of public interest that nobody properly reviewed, deepfake images, audio or video, and chatbots. A product description you drafted with ChatGPT is usually not one of them. An unreviewed article about new tenancy rules usually is.

This guide walks through the test the way a site owner would actually apply it, piece of content by piece of content. It is based on the regulation itself and on the Commission’s Article 50 guidelines and FAQ, not on vendor marketing.

First, which role are you in?

Article 50 splits the work between two groups, and most confusion comes from mixing them up.

  • Providers build the AI system: OpenAI, Google, Midjourney, the company behind your chatbot widget. They must make chatbots identify themselves (Article 50(1)) and embed machine-readable marks in generated output (Article 50(2)).
  • Deployers use an AI system in a professional context. If you run a business website and use AI to write, illustrate or answer visitors, you are a deployer. Your duties sit in Article 50(4): disclose deepfakes, and disclose AI-generated text published to inform the public on matters of public interest.

One consequence: the date you may have seen in the news, 2 December 2026, is not your deadline. The Digital Omnibus gave providers of generative AI systems already on the market before 2 August 2026 a grace period until then for machine-readable marking. Deployer obligations got no such delay. If you publish AI content today, the rules below already apply.

Text: the two-question test

For written content, ask two questions in order.

1. Is the text published to inform the public on a matter of public interest?

The Commission’s FAQ lists the kind of subjects it means: politics and democratic processes, public administration, justice, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific or cultural developments that could be the subject of public debate.

Read that list against your own site:

  • A WooCommerce shop’s product descriptions: generally not informing the public on a matter of public interest. No Article 50 label needed, although consumer law still forbids misleading claims.
  • A marketing agency’s blog post on “10 Instagram hooks for 2027”: not public interest in this sense.
  • A law firm’s article explaining new rent caps, a pharmacy’s post on vaccine side effects, an accountant’s summary of a tax change, a blog on EU regulation (like this one): yes, these inform the public on matters of public interest.

If the answer is no, Article 50 does not require a label for that text. You can still choose to add one, which is a trust decision, not a legal one.

2. Did a person with real expertise review it and take editorial responsibility?

Public-interest text is exempt when it has gone through human review or editorial control and a natural or legal person holds editorial responsibility for publishing it. Both halves matter, and the Commission has made the bar specific:

  • Human review means someone with relevant knowledge examined the substance: the facts, the sources, the conclusions. Spell-checking, grammar fixes or a quick skim do not count.
  • Editorial control means a person in practice has the authority to approve, change or reject the text on substantive grounds, including fact-checking.
  • Editorial responsibility means a named person or company carries the final legal responsibility for publication, including for the review process itself.

A byline alone is not review. Practitioners reading the guidelines also point out a trap: if a reviewer approves a draft and AI then substantially rewrites it, the approval no longer covers the new version. Re-review after heavy AI edits.

So for a public-interest article, you have exactly two lawful options: label it, or review it properly and keep a record that you did.

Images, video and audio: only deepfakes need a label

Article 50(4) does not require you to label every AI image. It requires disclosure of deepfakes: AI-generated or manipulated image, audio or video that resembles existing people, objects, places, entities or events and would falsely appear authentic or truthful. All three conditions must be met.

Image on your siteDeepfake?Why
Stylised illustration of a robot for a blog headerNoNobody would take it for a real photo.
Photorealistic “team photo” of people who don’t work for youLikely yesAppears to show real people at a real company.
AI photo of your real hotel with a pool it doesn’t haveYesResembles an existing place and would mislead.
Product photo with AI-removed backgroundUsually noStandard editing that doesn’t change what the product is.
Realistic image of a news event that didn’t happenYesResembles an event and falsely appears authentic.

For artistic, satirical or fictional work, the duty shrinks to disclosing that generated content exists, in a way that doesn’t spoil the work. And note what the Commission says about deployers relying on the provider’s hidden watermark: that is not enough. The disclosure has to be something a visitor can see or hear without special tools.

Chatbots: usually the provider’s job, still your problem

Article 50(1) puts the duty on the provider to design chatbots so people know they are talking to an AI, unless that is obvious from context. In practice, many widgets on WordPress sites (Tidio, Intercom, Crisp and others) leave the greeting text to you. If your bot opens with “Hi, I’m Anna from support” and never says it is an AI, a visitor has no way to know, and you are the one who configured it. Add a clear line at the start of the conversation. Our guide on chatbot compliance under the EU AI Act covers the wording and placement.

How a disclosure has to look

Article 50(5) sets three requirements for any label you do need:

  1. Clear and distinguishable. Not buried in your terms, not grey 9px text in the footer.
  2. At the latest at first exposure. Before or as the visitor sees the content, not on a separate page they may never open.
  3. Accessible. Screen readers must pick it up, so a label baked into an image with no text alternative fails.

The Commission has published an official set of EU icons that deployers may use for labelling, and a Code of Practice on transparency of AI-generated content that the Commission and AI Board have confirmed as an adequate way to show compliance. Signing the code is voluntary. If you don’t follow it, you need to be able to show your own approach is equally adequate. For placement patterns that work on real themes, see where to place AI disclosures on your website.

What happens if you get it wrong

Breaches of Article 50 can be fined up to 15 million euros or 3% of worldwide annual turnover, whichever is higher, with proportionality for SMEs. National market surveillance authorities enforce it. Realistically, a small site is unlikely to be the first target. The bigger near-term risk is a complaint from a competitor or a journalist pointing out an unlabelled deepfake. We break down the penalty structure in EU AI Act fines and penalties explained.

A 20-minute check for your own site

  1. List where AI touches your content. Blog drafting, product copy, images, translations, the chatbot, AI-powered search. Our website AI audit guide has a template.
  2. Sort your text into public interest or not. Most shop and marketing pages fall out here.
  3. For public-interest text, decide per piece: label or review. If you review, write down who checked what, against which sources, and when.
  4. Go through your images with the three-part deepfake test. Replace or label anything photoreal that depicts real people, places or events that never happened.
  5. Read your chatbot’s first message as a visitor would. If it doesn’t say it’s an AI, fix that today.
  6. Keep the record. If an authority asks, “we thought about it” is weaker than a dated list of decisions.

Doing this in WordPress without spreadsheets

If you publish a lot, tracking these decisions by hand gets messy. The free EU AI Act Ready plugin adds a disclosure level to every post (no AI, AI-assisted, AI-generated, or AI-generated and human-reviewed), shows a visitor notice only where one is needed, labels AI images, adds chatbot notices for 15 common widgets, and exports a report of what you decided. It runs on your own server and needs no account. More guides are on the EU AI Act Ready homepage.

Frequently asked questions

Do I have to label a blog post if I only used AI to fix grammar?

No. Assistive standard editing that doesn’t substantially change the text is outside the marking duty, and grammar fixes don’t make a post “AI-generated”. The label question only comes up when AI produced or substantially rewrote the substance.

Is an “AI-assisted” label legally required?

No. Article 50 has no “AI-assisted” category. It’s a useful transparency choice for readers, but the legal duties are the ones above: unreviewed public-interest text, deepfakes and chatbots.

Does Article 50 apply to sites outside the EU?

The AI Act applies to deployers established in the EU, and also to providers and deployers outside the EU where the output of the AI system is used in the EU. A US site aimed at EU visitors should assume it is in scope.

Our editor reviews every AI draft. Do we still need a label?

Not for the text, if the review is substantive, done by someone qualified, and your company takes editorial responsibility. Keep evidence of the review. Deepfake images in the same article still need their own disclosure.

Practical guide: How to label AI-generated images in WordPress.