
EU AI Act transparency rules for websites: chatbots, AI content and labels explained
What your site has to disclose:
Since 2 August 2026, the EU AI Act has had a short article that touches far more websites than its high-risk rules ever will. Article 50 decides when you have to tell people they are talking to an AI, when AI-made images, audio, video or text need a visible label, and who has to hide a machine-readable mark inside generated files. If your site has a chatbot, uses AI product photos, publishes AI-assisted articles or sells to people in the EU from anywhere else, some of it applies to you. This guide walks through the rules feature by feature, using the European Commission's own guidelines, its Code of Practice on AI-generated content and the EU's new label icons, and ends with a checklist you can work through in an afternoon. It explains the rules in plain language; it is not legal advice, so check anything high-stakes with a lawyer.
The short answer
For most business websites, Article 50 comes down to three habits. Say clearly, from the first message, that your chat assistant is an AI. Label AI images, audio and video that show real-looking people, places or events which did not happen as shown. And keep a real human editor in charge of anything you publish to inform the public, which takes most AI-assisted articles out of the labelling duty. Machine-readable watermarking is mainly a job for the companies that build the generators, but you should not strip it out.
The rules apply to businesses outside the EU too, as long as the AI output is used in the EU. Fines can reach 15 million euros or 3% of worldwide turnover, with proportionality promised for smaller companies.
First, work out which role you play
The AI Act splits Article 50 duties between two kinds of business, and most of the confusion about it comes from mixing them up.
A provider is whoever develops an AI system, or has it developed, and offers it under its own name or trademark. A deployer is whoever uses an AI system in a professional capacity. The Commission's FAQ adds that a provider is covered "irrespective of whether those providers are established or located within the EU or in a third country", and that freelancers who earn regularly from AI work count as deployers, not private individuals.
A typical small business website is a deployer of the image and text generators it uses. It can also become a provider without noticing: if you build a branded assistant on top of a model API and put it on your site under your own name, you have arguably put your own AI system into service. That matters, because the chatbot disclosure duty sits with providers. If you instead embed a ready-made widget from a vendor, that vendor is the provider, but your site is where visitors actually see the disclosure, so you still need to check it is there.
Your website features, mapped to the rules
Article 50 contains four duties plus a rule on how information must be shown. This is how they land on the features small business sites actually have.
| Website feature | Article 50 duty | Who carries it | What it means in practice |
|---|---|---|---|
| Chat assistant, AI agent or talking avatar | 50(1): tell people they are interacting with AI, unless obvious | Provider of the assistant | A clear notice at the start of the first interaction |
| AI image, video or audio generator offered to your users | 50(2): machine-readable marking and detection | Provider of the generator | Outputs carry metadata and watermarks; grace period to 2 Dec 2026 for systems already on the market |
| Emotion recognition or biometric categorisation (for example a camera-based mood tool) | 50(3): inform the people exposed | Deployer | Tell people before the system processes them |
| Realistic AI images, audio or video of people, places or events | 50(4): disclose deepfakes | Deployer (you, when you publish) | A visible or audible label at first exposure |
| AI-written articles on matters of public interest | 50(4): label, unless a human reviewed it and someone holds editorial responsibility | Deployer | An editorial process, or a label |
| All of the above | 50(5): clear, distinguishable, accessible, at first interaction or exposure at the latest | Both | No disclosures hidden in footers or terms pages |
Ordinary AI features that never talk to visitors directly, such as a recommendation engine, a search ranking model or an automated spam filter, sit outside Article 50(1). The Commission says systems "operating solely in the background, through machine-to-machine communication, or without direct contact with people, fall outside the scope of this obligation."
Chatbots: disclosure from the first message
The Commission's guidelines list four conditions that must all be met for the chatbot rule to apply: the system is an AI system, it is designed for a genuine two-way exchange, the AI itself communicates rather than a human relaying its output, and it talks to real people, whether consumers or professionals. A website support bot, a booking assistant or a sales agent that replies to typed or spoken questions meets all four.
Timing is strict. People must be told "from the start of the first interaction in a clear and distinguishable manner and in accordance with accessibility requirements." A line buried in your privacy policy does not do that, and neither does a disclosure that appears only after the visitor has asked three questions.
There is an exception when it is obvious that someone is talking to an AI, judged from the point of view of a reasonably well-informed, observant person. The Commission says the exception "should be interpreted in a restrictive manner", so a chat window that merely has a robot icon is a weak basis to rely on. The guidelines do name some narrow cases outside the rule, according to Il Sole 24 Ore's summary of the published text, such as coding assistants used only by professional developers, systems used exclusively by healthcare professionals and appliances with limited functions. A public website chat widget is not one of them.
Voice agents need the same treatment out loud: the first thing the caller hears should make clear that they are speaking with an automated assistant. If your chatbot hands over to a human, say so when it happens, so visitors know which replies came from a person. Our guide to adding an AI chatbot to your website covers the build side, and the one on protecting chatbots from prompt injection covers the security side.
<div class="chat-panel" role="dialog" aria-labelledby="chat-title">
<h2 id="chat-title">Ask our assistant</h2>
<p class="chat-notice" role="note">
You are chatting with an AI assistant, not a person.
It can make mistakes. Type "human" to reach our team.
</p>
<!-- messages render below the notice, so it is seen first -->
</div>AI images, video and audio: when a label is required
Not every AI image needs a label. The visible-labelling duty is about deepfakes, which the AI Act defines as AI-generated or manipulated image, audio or video content that "resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful." The Commission breaks that into three tests that must all be met: close resemblance, a subject that exists or plausibly could, and a real chance of misleading people about authenticity.
Context counts. The FAQ says deployers may weigh the level of resemblance, the message of the content, where it is shown and what the audience expects. An obviously illustrated blog header or a stylised product mock-up is unlikely to qualify. A photorealistic "customer" in a testimonial, a cloned voice of a real person, or a photo of your shop showing a refurbishment that has not happened can.
The EU's own examples are surprisingly close to everyday website work. On its icon page the Commission lists "authentic photographs of an empty apartment" that are "furnished using AI" as a case for the partially AI-modified label, which is exactly what virtual staging on property sites does. Swapping a face in a real photograph is another listed example.
When a label is needed, it must reach people at their first exposure, and the FAQ is explicit that deployers "cannot simply rely on the machine-readable marking embedded in the content by the provider". In other words, invisible watermarks from your generator do not count as telling your visitors. For evidently artistic, satirical or fictional work, the duty is lighter: disclose in a way that does not spoil the work, for example in credits or a caption next to it.
The EU icons, and how to place them
To make labels consistent, the EU published a free set of icons built around a capitalised "AI" mark, in three types: a basic icon, a fully AI-generated icon and a partially AI-modified icon, each in black, white and two 50% transparent versions. Using them is optional, but the Commission is blunt that "the labelling requirements under Article 50 AI Act are not" optional, and that using an icon does not prove compliance by itself.
The icons were user-tested, and the Commission reports that "performance improved across all measures when the basic icon was accompanied by a text label", so an icon with a word such as "modified" beats an icon alone. Its summary of the placement rules is a useful checklist even if you never sign the Code of Practice:
- Show the label clearly at the latest when a person first sees or hears the content.
- Do not put it where other interface elements can cover it.
- Embed it in the image, video or text itself where possible, so it survives when the content is shared or downloaded. An interface overlay is an acceptable alternative.
- Use plain language, a readable size, and alt text or an ARIA label so screen readers announce it.
- If a disclosure is shown only briefly, keep it on screen long enough to read. For audio-only content, Bird & Bird's analysis of the Code notes an audible disclaimer at the start.
AI-written text: the human-review route
The text rule is narrower than many people expect. It covers AI-generated or manipulated text that is published to inform the public on matters of public interest. The Commission's list of such matters includes politics, public services, justice, fundamental rights, security, public health, the environment, consumer safety and economic, financial, political, scientific or cultural developments that may be the subject of public debate. Product descriptions, booking confirmations and most marketing copy are not obviously in that zone; a news-style article about a new regulation or a health trend can be.
Even then, there is an exit that suits real businesses: no label is needed when the text has gone through human review or editorial control and a person or company holds editorial responsibility. The Commission sets a real bar. It says "superficial, solely formal, or procedural checks (e.g. spell-checking or grammatical correction) are not considered to be human review or editorial control." Review means someone with relevant knowledge checks the substance, including facts and sources, and can change or reject it.
In practice, write down who reviews AI-assisted articles, what they check and who signs off, and put a named author on the page. That is also simply good publishing: we describe our own process in our editorial policy. If you run fully automated pages, such as AI news summaries with no human in the loop, the EU's examples put those squarely in the fully AI-generated label category.
Machine-readable marking: what you need from your tools
Article 50(2) is the technical duty, and it belongs to providers of generative systems, including the large model companies. Their outputs have to be marked in a machine-readable way and be detectable as AI-made. Small sites mostly meet this rule as customers, not as builders, but there are two cases where it reaches you: if you offer visitors a generator of your own, such as an AI logo maker or image editor, or if you resell one under your brand.
The Code of Practice gives providers a recipe. Bird & Bird's summary of the final text says signatories mark outputs in at least two machine-readable layers: digitally signed, time-stamped metadata plus an imperceptible watermark, including for free-form text longer than 200 tokens, with fingerprinting or logging as an optional extra. They must also offer a detection tool, keep existing marks intact and not offer tools that remove them, and have a watermark-detection interoperability solution by 2 February 2027. Content Credentials based on the C2PA standard are the best-known form of the signed-metadata layer.
For everyone else the practical steps are modest: prefer generators that state they mark their outputs, avoid image pipelines that strip metadata by default when you do not need them to, and never use tools or settings designed to remove watermarks from AI content. The grace period here is short. Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, generative systems already on the market before 2 August 2026 have until 2 December 2026 to comply with the marking duty; anything launched since then had to comply from day one.
Dates, delays and who enforces it
A lot of the coverage of the Digital Omnibus talked about delays, and some of it was read as a delay to everything. It was not. The European Parliament's summary of the deal says it pushes high-risk obligations to 2 December 2027 and 2 August 2028 and delays the watermarking duty to 2 December 2026. The chatbot, deepfake and public-interest text rules still applied from 2 August 2026.
The Code of Practice was published on 10 June 2026. On 8 July the Commission concluded it adequately covers Articles 50(2), (4) and (5), and the AI Board followed the next day; by the end of July about 190 organisations had signed. The Commission's guidelines on Article 50 were approved in July and published in the EU's Official Journal on 12 November 2026. Launching the Code, Executive Vice-President Henna Virkkunen said: "Europeans have a right to know whether what they see, hear or read has been made or altered by AI, especially when such content can shape public debate. Transparency is how we protect trust."
Enforcement sits mainly with national market surveillance authorities, with the AI Office handling systems built on general-purpose models by the same company and AI inside very large platforms. Member states are still filling in the map. Denmark, for example, published a draft order on 2 November 2026 naming its Agency for Digital Government as the authority for Article 50, due to take effect on 1 December 2026 if the new Danish AI law is in force by then. Published enforcement cases are scarce so far: one tracker, Easy AI Act Labels, found no known fine or court ruling over AI image labels in the EU as of 30 September 2026. That is not a reason to wait; maximum fines are 15 million euros or 3% of worldwide annual turnover.
Selling into the EU from outside it
The AI Act reaches providers and deployers based outside the EU when the output of their AI system is used in the EU. For a website, that is a low bar: a chatbot that answers visitors in Germany or France, or AI-made product photos shown to shoppers in Spain, is output used in the EU.
So a store in Pakistan, the US or the UK with EU customers should treat Article 50 as part of selling there, in the same way it treats EU consumer law or GDPR. The cheapest approach is to apply the same disclosures worldwide rather than geo-targeting them: the notices are short, visitors increasingly expect them, and platforms such as Meta, Google and TikTok already label AI content in their own ways, as we covered in our guide to AI video ads.
An afternoon checklist for your site
Bird & Bird's Simon Hembt and Oliver Belitz sum up the position for companies that are unsure whether to sign the Code: "Either way the work is unavoidable, because signed or not, a company needs an internal compliance regime." For a small website, that regime can be short:
- List every place AI touches your site: chat or voice assistants, generated images and video, AI-written or translated pages, tools you offer visitors, and any camera or emotion features.
- For each item, decide whether you are the provider (it runs under your name) or the deployer (you use someone else's tool to publish).
- Add a plain AI notice to every chat or voice assistant, shown before or with the first reply, announced to screen readers, and repeated on handover to a human.
- Go through your images, video and audio. Label anything realistic that shows people, places or events that did not exist or did not happen as shown, especially staged rooms, edited faces and synthetic voices. Use the EU icon with a short text label if you like.
- Write a one-page editorial process for AI-assisted articles: who checks facts and sources, who approves, and the named author. Label any fully automated public-interest text you keep.
- Check your generators and image pipeline: do the tools mark their outputs, and does your CMS or compression step strip that data needlessly?
- Record what you decided and why, with the date. If a regulator or a customer asks, a short written record is far better than a memory.
FAQ
Does the EU AI Act apply to my website if my business is outside the EU?
It can. The Act covers providers and deployers outside the EU when the output of their AI system is used in the EU. A chatbot answering EU visitors or AI images shown to EU shoppers is output used in the EU.
Do I need to label every AI-generated image on my site?
No. The visible-label duty covers deepfakes: realistic AI-made or AI-altered images, audio or video of existing or plausible people, places, objects or events that could be mistaken for real. Clearly illustrative graphics usually do not qualify. Labelling more than the minimum is allowed and often builds trust.
Do AI-assisted blog posts need an AI label?
Only text published to inform the public on matters of public interest is covered, and even then no label is needed if a knowledgeable person reviewed the substance and someone holds editorial responsibility. Spell-checking alone does not count as review.
When did the chatbot disclosure rule start?
On 2 August 2026. The Digital Omnibus only delayed the machine-readable marking duty for generative systems already on the market, to 2 December 2026.
Do I have to use the EU AI icons?
No. The icons are optional and free to use without attribution. The duty to disclose is not optional, and the Commission says using an icon does not prove compliance on its own.
Sources
- European Commission: Transparency obligations under Article 50 of the AI Act (FAQ)
- European Commission: Quick Facts, transparency rules for AI systems
- European Commission: Guidelines on transparency obligations for providers and deployers of certain AI systems
- European Commission: Code of Practice on Transparency of AI-generated Content
- European Commission press release: Code of Practice on marking and labelling AI-generated content (IP/26/1328, Jun 10, 2026)
- European Commission: Opinion on the assessment of the Code of Practice on Transparency of AI-generated Content (Jul 9, 2026)
- European Commission: EU icons for labelling AI-generated content
- European Parliament: AI Act, deal on simplification measures, ban on nudifier apps (May 7, 2026)
- EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)
- Il Sole 24 Ore: AI, the transparency requirement does not apply to developers (Marina Castellaneta, Nov 21, 2026)
- Bird & Bird: Taking the EU AI Act to practice, the final Transparency Code of Practice (Simon Hembt, Oliver Belitz, Jun 22, 2026)
- Lovguiden: Draft Danish order naming the Agency for Digital Government as Article 50 authority (Nov 2, 2026, in Danish)
- Easy AI Act Labels: AI labelling, who enforces it and which cases are known (as of October 2026)
- C2PA: Implementation guidance (specification 2.4)





