Since 2 August 2026, Article 50 of the EU AI Act has been in force. It is the part of the Act that touches almost every small business using AI, because it covers the things SMBs actually do: website chatbots, AI phone receptionists, AI-generated images and video, and AI-written content. The Digital Omnibus postponed the high-risk rules, but it did not postpone these transparency duties.
The advice circulating online swings between two extremes: "label everything AI touched" and "this only applies to big tech". Neither is right. This guide sets out what you actually have to disclose, what you do not, and how to do it in an afternoon. It is practical guidance, not legal advice. If a use case is borderline, check it with a lawyer who knows the Act.
What Article 50 actually requires
Article 50 splits duties between two roles. Knowing which role you are in matters more than any labelling template.
- Providers build or supply the AI system: OpenAI, Anthropic, Google, Microsoft, your chatbot vendor. They must design chatbots so people know they are talking to AI. They must also mark generated images, audio, video and text in a machine-readable way, such as watermarks or metadata.
- Deployers use an AI system in their business. That is you, most of the time. Deployers must disclose deepfakes, disclose AI-generated text published to inform the public on matters of public interest, and tell people when emotion recognition or biometric categorisation is used on them.
There is one grey area to watch. If you build your own chatbot or voice agent on top of a model API and launch it under your brand, you may count as the provider of that system as well as its deployer. In practice, treat chatbot disclosure as your job whoever built the bot.
Fines for breaching Article 50 go up to €15 million or 3% of worldwide annual turnover. For SMEs the lower of the two applies. Nobody expects a 10-person firm to face that ceiling, but national regulators can now enforce, and a missing disclosure is one of the easiest breaches to prove.
The five situations that trigger a disclosure
For a typical small business, these five uses create real obligations.
1. A chatbot on your website, WhatsApp or app
People must be told they are interacting with AI unless it is obvious to a reasonably well-informed person. Do not rely on "obvious". A widget called "Sophie" with a smiling headshot is not obvious. Put the disclosure in the first message, not in your privacy policy.
2. An AI voice agent answering your phone
The same rule applies, and the risk of confusion is higher because modern voices sound human. Disclose in the first sentence of the call: "Hi, you've reached Martinez Plumbing. I'm the AI assistant. I can book a visit or pass you to the team." Our guide to AI voice agents for small business covers set-up and cost.
3. Realistic AI images, video or audio of people, places or events
This is the deepfake rule. It covers AI-generated or manipulated content that resembles real people, objects, places or events and would falsely appear authentic. For an SMB, typical examples are:
- A photorealistic "customer" holding your product in an advert
- An AI video of your founder delivering a message they never recorded
- A cloned voice of a staff member in a radio or social ad
- A "before and after" photo of a renovation that was generated, not photographed
Creative, satirical or clearly fictional work gets a lighter form of disclosure. Advertising and product imagery do not count as artistic, so they need a clear label.
4. AI-written text on matters of public interest
If you publish AI-generated text to inform the public on matters of public interest, you must disclose it. There is an important exception: content that a person has reviewed, where an identified person or organisation takes editorial responsibility. More on that below.
5. Emotion recognition or biometric categorisation
This is rare in small firms, but it does turn up in some call-analytics and retail-footfall tools. If a system infers customers' emotions from their voice or face, you must tell them. Using emotion recognition on employees is prohibited outright, except for medical or safety reasons. Analysing the sentiment of written reviews is not emotion recognition in the Act's sense.
What you do not have to label
Over-labelling has costs. It clutters your pages and teaches customers to ignore the label. You generally do not need an Article 50 disclosure for:
- Blog posts, newsletters and web copy drafted with AI and edited by you. Most marketing content is not a "matter of public interest". Even when it is, the human-review and editorial-responsibility exception applies if a named person or your business stands behind it.
- Emails, proposals and quotes written with AI help and sent by a person.
- Obviously artificial visuals, such as stylised illustrations, icons, cartoon mascots or abstract backgrounds that nobody would take for a photograph.
- Minor edits, such as AI background removal, colour correction, upscaling or spelling fixes that do not change what an image or text shows.
- Internal use: meeting notes, summaries, analysis and drafts that never reach the public.
Not legally required does not mean irrelevant. Clients may still expect to be told, especially in professional services. That is a trust decision rather than a compliance one, and we cover it in should you tell clients you use AI?
A quick test for any piece of content
- Could someone reasonably believe a real person, place or event is shown when it is not? → Label it.
- Is someone talking to an AI without knowing it? → Disclose it.
- Is it published text that informs the public on a matter of public interest, with no human review and no named person accountable? → Label it, or add review and a named editor.
- None of the above? → No Article 50 label needed. Your own transparency policy decides.
How to label: wording and placement that works
The Act requires disclosure that is clear and distinguishable, given at the latest at the first interaction or exposure, and accessible. It does not prescribe exact wording. The Commission-endorsed Code of Practice on transparency of AI-generated content is voluntary, but it is the best available signal of what regulators will regard as adequate. These formats are short and work well:
- Website chatbot: "I'm an AI assistant. Type 'human' at any time to reach our team." Placement: First message, plus a small "AI" badge on the widget.
- Phone voice agent: "I'm [Business]'s AI assistant." Placement: First sentence of every call.
- Realistic AI image: "AI-generated image" Placement: On or directly under the image, not only in the alt text.
- AI video or cloned voice: "This video uses AI-generated footage/voice" Placement: On screen at the start, and in the caption.
- Unreviewed public-interest text: "This article was generated with AI." Placement: Top of the article, not the footer.
Two practical points. First, do not strip the metadata or watermarks your tools add to generated files. Providers had to start marking output from 2 August 2026, with a transition to 2 December 2026 for systems already on the market. Removing those marks makes your disclosure harder to defend. Second, make labels accessible: a screen reader should announce them, and they should not disappear on mobile.
If you generate a lot of marketing imagery, our guide on using AI image generation for small business marketing shows where realistic images help and where an obviously illustrated style avoids the question entirely.
A 60-minute transparency audit
You do not need a compliance platform. Block out an hour and work through this.
- List every customer-facing AI touchpoint (15 minutes). Chat widgets, phone lines, booking assistants, social replies, ad creative, product photos, video, podcast voice. Ask the team what they use, because shadow AI is where gaps hide.
- Run each through the four-question test above (15 minutes). Mark each one: label, disclose, or no action.
- Fix the live gaps (20 minutes). Edit the chatbot's opening message, the voice agent's greeting and the captions on realistic images. Most tools let you change these in settings.
- Write a five-line rule for the team (10 minutes). For example: "Realistic AI people or places always get an 'AI-generated' label. Chatbots and voice agents always say they are AI. Every published article has a named reviewer. We never remove AI watermarks." Add it to your AI policy.
Repeat the audit whenever you add a new customer-facing tool, and at least once a year.
The bottom line
For most small businesses, Article 50 comes down to three habits. AI that talks to customers says it is AI. Realistic synthetic people, places and events get a visible label. Published content has a human who reviewed it and stands behind it. Get those right and you have covered the transparency rules without plastering "AI" over everything you publish.
For the rest of the Act, including risk tiers, the AI literacy duty and the postponed high-risk deadlines, see our EU AI Act compliance guide for small businesses.
Where does your business stand on AI?
Take the free 3-minute AI Readiness Quiz and get a personalised score with your next steps.
Take the Free Quiz →