Back to the blog

AI Act Article 50 in practice: chatbot notices, deepfake labels and human review

· By Bee Flow · 6 min read

Since 2 August 2026 Article 50 of the AI Act applies. What it asks of an ordinary company, with the Commission's own examples: the chatbot notice, deepfakes in marketing, when AI text needs a label and what counts as review by a person.

A speech bubble, a framed picture and a printed page, each with a small amber tag tied to it

Since Sunday 2 August 2026 the transparency rules in Article 50 of the AI Act apply. The amendment that entered into force at the end of July, the Digital Omnibus on AI (Regulation (EU) 2026/1744), left that start date in place, except on one point for makers of generative AI. On 20 July the European Commission published guidelines on Article 50 full of examples. Below we go through them for anyone running a website, chatbot or content calendar. It is an explanation, not legal advice.

Four duties, two roles

An AI system that talks directly with people must let them know it is AI (paragraph 1). Systems that generate images, audio, video or text must mark that output in a machine-readable way (paragraph 2). Anyone using emotion recognition or biometric categorisation must inform the people exposed to it (paragraph 3). And anyone publishing a deepfake or certain AI-generated text must say so (paragraph 4).

Paragraphs 1 and 2 fall on the provider, who places the system on the market or puts it into service under its own name. Paragraphs 3 and 4 fall on the deployer: the company that uses the system in its work. For an ordinary small company that mostly means paragraph 4, and the chatbot on its own website.

The chatbot on your website

According to the Commission, a company that builds a chatbot itself and puts it into service under its own name is a provider. If you buy a ready-made widget, the duty sits with the vendor, but the notice still appears on your site.

A mention only in the terms and conditions, a vague word like "assistant" or a general line such as "Services on this website use AI" is not enough. What works, per the guidelines: a plain sentence like "You are interacting with an AI system" in the chatbot's first turn, optionally with a persistent label and a short note close to the input field. A human profile picture next to the bot can mislead.

One clear notice before the first exchange is usually enough. For financial, legal or health advice and for complaints handling, the Commission considers reminders during the conversation likely to be necessary. And when a visitor asks whether they are talking to a person, the system has to say it is AI.

The "it is obviously AI" exception is to be read narrowly. The Commission gives an internal assistant for trained staff who know they are using AI as an example where no notice is needed. A chatbot open to the public does not qualify.

Paragraph 1 does not cover a customer service employee who drafts a reply with AI and sends it themselves: the customer is not interacting with the system directly. The same goes for an e-mail a colleague wrote with AI's help. Where a service mixes AI answers with human ones, the AI parts must be disclosed unless a person properly reviewed them and sent them.

Deepfakes in your marketing

A deepfake is AI-generated or manipulated image, audio or video that resembles existing people, objects, places or events and would falsely appear authentic. A video with a realistic AI avatar of the CEO thanking staff for the past year is a deepfake. So is an AI product image in an advert or on packaging that makes the product look better or different from the real thing. An advert with talking mice arguing about cheese is not, because nobody takes it for real.

The label has to be visible or audible. The invisible watermark a vendor's tool adds does not count, the Commission writes. Evidently artistic, satirical or fictional work only needs a lighter disclosure. The Commission does not count a teleshopping-style advert with simulated customers as creative work.

A printed page under a magnifying glass, with a fountain pen beside it and an amber check mark in the margin

When AI text needs a label

This covers only text published with the purpose of informing the public on matters of public interest. Published means accessible to a large, indeterminate audience. An internal memo, an intranet post, professional correspondence between people and a draft in your folder are not published. Public interest covers topics such as politics, health, the environment, consumer safety and economic developments that merit public debate.

Advertising copy and product descriptions fall outside, according to the Commission, unless they make claims about, for example, health, consumer safety or sustainability. An AI-edited report with investor information on a listed company's website falls inside.

What counts as reviewed by a person

AI text on matters of public interest needs no label when two conditions are both met. A person has reviewed the substance of the text or it went through editorial control. And a person or organisation holds editorial responsibility for the publication.

Human review, per the Commission, means someone with knowledge of the subject deliberately examining the content, with fact-checking as the minimum. A spell-check, an editorial policy that exists only on paper, an automated check or a cursory sign-off is not enough. If AI rewrites the text after sign-off, the exception falls away.

The dates and the one transition period

The only transition period is in the new Article 111(4). Providers of generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking of paragraph 2. There is no delay for the chatbot notice or for the labels in paragraph 4. Deepfakes generated before 2 August and text published before that date do not need labelling after the fact, according to the guidelines. AI text you made earlier and publish now does.

The voluntary Code of Practice on Transparency of AI-Generated Content, final since 10 June, has a section on labelling deepfakes and text. Breaches can lead to fines under Article 99 of up to 15 million euros or 3 percent of worldwide annual turnover, and for SMEs the lower of the two applies.

What to sort out this month

  • Open your own chatbot as a visitor and read its first line. If it does not say in plain words that it is AI, ask the vendor for an opening line or add one yourself.
  • Add two columns to your content calendar: AI used, and reviewed by. Whoever reviews checks the facts and adds their name.
  • Agree that nobody runs a text through AI again after sign-off.
  • Decide who holds editorial responsibility for the blog and newsletter, and name that role where people can find it, for example in the legal information on your website.
  • Settle on one standard label for images and video, such as "This video was made with AI", for avatars, cloned voices and AI product shots.
  • Put these agreements in your one-page AI policy.

For the other duties of the regulation, start with the AI Act for small companies.

Where Bee Flow helps

Bee Flow's Compliance Hub checks whether each published agent says it is AI, and can add a sentence with that notice to the instructions of agents that lack one. Automations can include an approval step, so an AI draft only moves on once a colleague has read it. Answers from a knowledge base link to the source document, which makes fact-checking easier. How we handle privacy and security is on Privacy & Security. To see what this looks like for your team, get in touch.

#compliance #ai-at-work