AI Act Article 50 Is Live: Small Businesses Outsourcing Content Need to Settle Who Labels What

Disclaimer: I’m not a lawyer, and this isn’t legal advice. It’s my reading of the European Commission’s guidelines and published legal analysis, with a practitioner’s view of what matters, meant to help small business owners know which questions to ask. Check your own situation with a lawyer.

This category is where I’ll occasionally write up regulatory changes that European business owners, and small ones in particular, would want to know about. The first post covers Article 50 of the AI Act, because it’s already in force and nearly every company using AI for content or customer service runs into it.

In force, and not postponed

Article 50 of the EU AI Act sets out a group of transparency obligations, and they have applied since 2 August 2026. The Digital Omnibus adopted this summer pushed the high-risk rules back to December 2027, but left Article 50’s main dates alone. The one grace period covers machine-readable watermarking, an obligation on tool providers: generative AI systems already on the market before 2 August have until 2 December 2026 to comply. For most small businesses, which use these tools rather than build them, that grace period doesn’t change much.

The Commission published guidelines on Article 50 on 20 July. They aren’t legally binding, but national authorities are likely to lean on them when enforcing. Breaches can draw fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher.

What needs a label, and what usually doesn’t

Realistic synthetic images, video and audio. The guidelines call these deepfakes: content generated or altered by AI that resembles real people, objects, places or events and would falsely appear authentic. The label has to be clearly visible or audible the first time someone encounters the content. Routine retouching and colour correction usually don’t need one; staging a scene that never happened does.

Cloned or lifelike voices in ads usually count too. An AI voiceover that imitates human intonation and breathing can need a label even when the footage itself is real. The Article treats evidently artistic, creative, satirical or fictional works more lightly, but that lighter treatment lets you disclose in a way that doesn’t spoil the work; it doesn’t remove the duty to disclose. The guidelines also say these categories will be read strictly, and where content mixes information and creativity, the informative side wins. An ad selling a product with an AI voice is hard to pass off as art.

Ordinary social copy and product descriptions are usually outside the text rule. Article 50’s text obligation targets AI-generated text published to inform the public on matters of public interest, such as politics, public health, justice, consumer safety, or financial and scientific developments. A product post or an item description generally isn’t that.

Be careful with articles that inform the public on tax, health or legal topics. If AI generates them, they need a label unless they’ve gone through substantive human review and a named person holds editorial responsibility. The guidelines stress that the review has to be substantive, not a glance, and that the exception falls away if AI makes material changes after sign-off. An accountancy firm that has AI write a weekly tax update and publishes it without anyone properly reviewing it is on the wrong side of that line.

Bots that answer customers on their own have to say they’re AI at the first interaction. That covers website chatbots, WhatsApp support, and any n8n or Make workflow you’ve built that replies to customer emails by itself. According to the guidelines, a vague label like “assistant”, a line in the terms and conditions, or a machine-readable marker alone won’t do. Strictly, this is an obligation on the system’s provider, and a company that assembles its own workflow to answer customers may well count as the provider. It’s the reason the chat assistant on my own site says, right above the input box, that you’re talking to an AI and not a person.

Where it most often goes wrong: who applies the label

Article 50’s labelling duty for content falls on the deployer, meaning whoever uses the AI tool under their own authority to produce and publish the content. Tool developers aren’t the target. Using a third-party tool such as Midjourney to produce a client’s posts and ads is enough to make you a deployer.

The trouble sits between brands and agencies. A guide published in September by IAB Austria and the law firm act legal argues that in most outsourced work it’s the agency that actually operates the AI, so the duty usually lands on the agency. Plenty of cases are murkier: the brand reworks the agency’s copy with AI, the agency subcontracts to a freelancer, or the brand supplies the assets and the agency only lays them out. The Commission’s guidelines likewise suggest reviewing contracts with creative agencies and content suppliers so that the labelling duty follows whoever actually makes the decisions.

The cheapest practical step is to write the answers into the brief and the contract:

  • which assets may use AI, and which may not
  • who decides whether something needs a label, and who applies it
  • who confirms whether content depicts real people, places or events
  • who keeps the evidence, such as screenshots before publication, review records in the CMS, and where the label appeared and what it said

A starting order for small businesses

If your company outsources social, advertising or website content, or uses AI for content and customer service itself, I’d take it in this order.

  1. Make a list of the tools. Who inside and outside the company uses which AI tools, and for what. Many companies only discover at this point that interns and contractors use more tools than anyone knew.
  2. Set rules for data that stays out. Which customer data, contracts and internal documents must never go into an AI tool. That’s mainly a GDPR and confidentiality question rather than an Article 50 one, but the two conversations tend to happen together.
  3. Then agree label templates. One wording and placement each for images, video and audio, such as “Image generated with AI” or “This voiceover uses a synthetic voice”, so the team doesn’t reinvent it every time.
  4. Put responsibility into the contract and the brief, along the lines of the list above.
  5. Check your customer-facing bot. Does its first message make clear that it’s an AI?

Most of this is tidying up and talking to people, and it costs little. I think it does more good than buying yet another “compliance chatbot”, because what Article 50 ultimately asks is that a company knows where it uses AI and that someone stands behind what it publishes.

Sources