EU AI Act Article 50: Transparency Rules Checklist for 2026
Article 50 of the EU AI Act applies since August 2, 2026. What it means for chatbots, AI images and AI text on your website, with dates, fines and a checklist.
Working through a similar problem? Tell us what you want to build.
Legacies is an independent digital product brand based in Bucharest, Romania, founded by Horia Stan and Alexandru Tălnaci, and since August we have been checking every AI feature we build against one short article of EU law. It touches ordinary things: a chatbot on a website, an AI-generated image in a campaign, a blog post written by a model.
Here is the core of it. Since August 2, 2026, Article 50 of the EU AI Act requires that people are told when they are talking to an AI system, that AI-generated content is marked, and that deepfakes and certain AI-written public-interest texts are labelled. It applies to businesses outside the EU too, whenever the AI system or its output reaches people in the EU.
Quick answer
- Article 50 applies from August 2, 2026. The Digital Omnibus delayed high-risk AI rules, but it did not delay these transparency rules.
- Chatbots must say they are AI. Unless it is obvious to a reasonably well-informed person, users must be told at the latest at their first interaction.
- Deepfakes and public-interest AI text need labels. Text is exempt only if a person reviews it and someone holds editorial responsibility.
- Fines reach EUR 15 million or 3% of global turnover. For SMEs and startups, the lower of the two applies.
What Article 50 actually requires
The AI Act sorts AI systems by risk. Most website tools are not "high-risk". They fall into the transparency category covered by Article 50, which splits duties between providers (who build or put an AI system on the market) and deployers (who use it professionally).
| Situation | Who is responsible | What is required |
|---|---|---|
| AI that talks directly with people (chatbots, voice agents) | Provider | Design it so people know they are dealing with AI, unless obvious |
| AI that generates audio, images, video or text | Provider | Mark outputs in a machine-readable, detectable way |
| Emotion recognition or biometric categorisation | Deployer | Inform the people exposed to it |
| Deepfakes (realistic image, audio or video) | Deployer | Disclose that the content is AI-generated or manipulated |
| AI text published to inform the public on matters of public interest | Deployer | Disclose it, unless a human reviewed it and holds editorial responsibility |
The information must be clear and distinguishable, and given at the latest at the first interaction or exposure. It also has to meet accessibility requirements.
Provider or deployer: which one are you?
Most small businesses are deployers. You add a chatbot from a vendor, generate images with a tool, publish text drafted by a model. The vendor is usually the provider and handles things like watermarking.
But the line can move. If you build your own assistant on top of a model API and put it on your website under your name, you may take on provider duties for that system. The Commission's guidelines on Article 50 include practical examples worth reading if you build rather than buy.
Does it apply outside the EU?
Yes, in many cases. The AI Act covers providers placing AI systems on the EU market wherever they are based, and providers and deployers outside the EU when the output is used in the EU. A US or UK company with a chatbot serving EU customers should treat Article 50 as relevant.
The dates that matter
The timeline moved this summer, so older articles may be wrong. Here is where things stand as of October 2026:
- June 10, 2026: the Commission published the final Code of Practice on Transparency of AI-Generated Content. It is voluntary and includes an optional EU icon for labels.
- July 20, 2026: final Commission guidelines on Article 50 published.
- July 27, 2026: the Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force. It pushed stand-alone high-risk obligations to December 2, 2027 and product-embedded ones to August 2, 2028.
- August 2, 2026: Article 50 transparency obligations apply.
- December 2, 2026: end of the grace period for machine-readable marking of outputs from generative AI systems already on the market before August 2, 2026. Systems launched after that date get no grace period.
The grace period covers the provider's marking obligation. It does not delay chatbot disclosure or deepfake labelling for deployers.
What it means for a typical business website
Chatbots and AI assistants
If a visitor opens a support chat and might assume a human will answer, tell them. The exception for cases where AI is "obvious" is narrow. The final guidelines say a general awareness that AI exists is not enough, and the bar is higher when users include children, older people or people with disabilities.
A clear line such as "You are chatting with our AI assistant" beats a friendly human name with no explanation. Even if the human-sounding version gets more replies at first, the trust cost when people find out is higher.
If the assistant can take actions, like booking an appointment or changing an order, make the difference between a suggestion and a confirmed outcome obvious. We cover how to scope those actions in why AI agents should start with one boring workflow.
AI-generated images, video and voice
Realistic generated media that could pass as real is a deepfake under the Act, even if you had no intent to deceive. A product photo with an AI-generated model, a voice-over cloned from a real person, a fake "customer" in a video: all need a visible label. Obviously fantastical illustrations are not deepfakes. Artistic, creative or satirical work gets a lighter regime: you flag that AI was used without spoiling the work.
AI-written blog posts and articles
The text rule is narrower than many people think. It covers AI-generated text published to inform the public on matters of public interest, such as politics, health, consumer safety, finance or science. A product description is usually not that. A news-style article about a new regulation might be.
The exemption is real review: a competent person edits the text and someone takes editorial responsibility for it. The guidelines set a high bar. A quick spell-check or a rubber-stamp sign-off does not count.
A practical checklist for your AI features
- List every AI touchpointChat widgets, automated emails, voice agents, generated images, AI-written pages. Include tools a vendor added for you.
- Label the interaction where it happensPut the disclosure next to the chat or in the first message, not only in your terms and conditions.
- Separate drafting from actingMake it clear what the system only suggests and what it actually confirms, books or changes.
- Define human reviewDecide who reviews public claims, prices and advice, and write it down.
- Ask vendors about markingCheck that your image, video and text tools mark outputs, especially if they launched after August 2, 2026.
- Keep a simple recordNote what each AI feature does, who owns it and when you last tested it as a first-time visitor.
Better interfaces, not more warnings
The businesses that handle this well will make AI feel less mysterious. A good assistant says what it knows, what it cannot do and when it is handing over to a person. That is a stronger experience than pretending every answer came from an employee.
This is also why a generic chat widget is rarely the final answer. A useful assistant needs your real context: current services, policies, availability, language and escalation paths. When we build a website or web app with AI in it, the work is not connecting a model to a chat bubble. It is deciding the job, the permissions, the handoff and the disclosure. If that is what you need, our services page lists custom web apps starting at 3,499 lei (about 665 EUR) and websites starting at 2,199 lei (about 415 EUR).
Transparency also connects to who reads your site. AI crawlers now summarize business websites, so your public facts need to be accurate. More on that in AI crawlers and robots.txt. And if your developers use AI to write code, the same principle of a human owning the result applies, as we explain in how AI coding agents are changing small teams.
When to handle it yourself and when to bring someone in
Handle it yourself if you use one or two off-the-shelf tools, like a chat widget or an image generator. Open your site as a first-time visitor, try every automated interaction, add clear labels and ask your vendors about marking. That covers most small sites.
Bring someone in if you built your own AI features, if your assistant can take actions on a customer's behalf, if you publish a lot of AI-assisted content on public-interest topics, or if your audience includes children. In those cases, you may need technical changes and legal advice. For a quick look at the rest of your site's health, run our free website audit.
Frequently Asked Questions
When does Article 50 of the EU AI Act apply?
Article 50 applies from August 2, 2026. The Digital Omnibus on AI, which entered into force on July 27, 2026, delayed high-risk obligations but not these transparency rules. The only extension is for machine-readable marking by generative AI systems already on the market before August 2, 2026, which have until December 2, 2026.
Do I have to tell website visitors they are talking to a chatbot?
Yes, unless it is obvious to a reasonably well-informed and observant person that they are dealing with AI. The Commission's guidelines treat that exception narrowly, so the safe approach is a clear statement in or next to the chat at the first interaction, such as saying the visitor is chatting with an AI assistant.
Do I need to label AI-generated images on my website?
You need a visible label if the image, audio or video is a deepfake, meaning realistic content that could plausibly pass as real, such as a generated person or a cloned voice. Clearly fantastical illustrations are not deepfakes. Separately, the tool provider must mark generated outputs in a machine-readable way.
Do AI-written blog posts need a disclosure under the EU AI Act?
Only if the text is published to inform the public on matters of public interest, such as politics, health, finance or consumer safety, and it was not properly reviewed. If a competent person edits it and someone takes editorial responsibility, the disclosure duty does not apply. Ordinary marketing and product copy usually falls outside this rule.
Does the EU AI Act apply to companies outside the EU?
Often, yes. It covers providers placing AI systems on the EU market wherever they are based, and providers and deployers outside the EU when their AI output is used in the EU. A US or UK business running a chatbot for EU customers should follow Article 50.
What are the fines for breaking the AI Act transparency rules?
Breaches of Article 50 can be fined up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs, including startups, the cap is whichever of the two amounts is lower. Enforcement is handled by national market surveillance authorities in each EU country.