skip to content

marketing 13 min read

Labelling AI content from 2 August 2026: what a small business has to do

Labelling AI content has been mandatory since 2 August 2026. Check which images, videos and chatbots in a small business need a label and how to do it without panic.

Labelling AI content from 2 August 2026: what a small business has to do

Article 50 of the EU AI Act has applied since 2 August 2026, and with it the labelling of AI content: chatbots have to say that they are a machine, deepfakes get a mandatory label, and generative tools have to mark files technically. For a small business it mostly means a few things to put in order. Below I explain calmly what exactly changes and where to start.

What the AI Act changes from 2 August 2026

What the AI Act changes from 2 August 2026

Image generated by AI

Article 50 of the AI Act, that is EU Regulation 2024/1689, applies from 2 August 2026. From that day the labelling of AI content covers specific duties to do with chatbots, deepfakes, selected texts and the technical marking of files.

The aim of the rules is simple: the recipient should know when they are talking to an AI system or looking at realistic material generated or altered by artificial intelligence.

Article 50 splits the duties between the providers of AI systems and the businesses that use them. The provider of an image, text, audio or video generator is responsible for putting a machine-readable mark in the file. It can be information saved in the metadata or a standard that confirms where the material came from.

The business publishing the content is responsible above all for a visible label on a deepfake. This duty covers realistic images, recordings and films that show an existing person, place, object or event in a way that could mislead the recipient. The new rules were covered in more detail by Gazeta Prawna (a Polish legal daily).

An ordinary illustration showing a fictional scene does not automatically become a deepfake. Article 50 of the AI Act does not order a business to caption every graphic made by a generator. Your own rule of labelling all such material is a sensible practice all the same. It makes publishing simpler and tells recipients clearly how the content was made.

The third area is chatbots. A system holding a conversation with a person should disclose that it is AI, unless that is clear from the context itself. The information has to appear at the latest on first contact.

The AI Act still lets you use text, image, voice and video generators. Article 50 brings in transparency rules for specific uses.

Who has to label AI content

Who has to label AI content

Image generated by AI

The duties apply to professional use of AI systems. The size of the business does not matter here, so the rules can cover a sole trader as much as a large company.

Use counts as professional when it is connected with a business, trade, a profession or freelance work. That includes online shops, agencies, influencers who earn from their posts, local businesses running social media profiles and people promoting their own services.

Private use stays outside the scope of the AI Act. An example would be a collage made for the family or a film sent to friends, as long as the person publishing it is not acting in connection with their work or business.

Different kinds of material made with the help of AI can appear in a company’s channels. Each one has to be assessed on its content and the way it is used. A generated illustration on a blog will usually not be a deepfake if it shows a fictional scene. A visible label can be added voluntarily.

It is different with an “animated” photo of a real product. If a film shows the product realistically but gives it features or behaviour it does not actually have, it may meet the definition of a deepfake. A clear label will be needed.

A similar assessment applies to a synthetic voice-over. A voice created from scratch that does not imitate a specific person is subject to technical marking on the tool provider’s side. A voice pretending to be the business owner, an employee or a well-known person may need a visible or audible label as a deepfake.

A video avatar also has to be assessed for its likeness to a real person and its potential to mislead the recipient. The more realistic the material and the closer its link to a specific person, place or event, the more likely it is that a label will be needed.

Small corrections made with the help of AI usually do not trigger the technical marking duty. That covers standard adjustments to sharpness, white balance or the crop, or removing a small, incidental element from the background. Replacing the whole background, adding new objects, changing what someone says or reworking the material in a way that changes its meaning goes beyond ordinary editing.

How to label AI images, video and audio

A deepfake has to be labelled clearly and in a way the recipient can access, at the latest on first contact with the material. For other graphics, films and recordings a business can adopt a broader labelling rule to stay consistent across all channels.

How to write the label

The rules do not impose a single wording. The message should let the recipient understand quickly that the material was created or altered using artificial intelligence.

You can use one of these simple labels:

  • “AI-generated”
  • “Modified by artificial intelligence”
  • “Illustration generated by AI”
  • “This recording contains an AI-generated voice”

With a deepfake it is worth describing the matter more precisely, especially when the material shows a specific person or event. An example message could be: “This material was generated by AI and is not a real recording of this event”.

Alongside the EU code of practice now being drafted, a set of ready-made icons for labelling AI content has also appeared. Using them can make it easier to keep a common standard across several language versions. The icon itself should be understandable in the given context. If in doubt, it is better to pair it with a short text description.

Where to put the information

A label hidden in the site’s terms and conditions or on a separate subpage will not meet the first-contact requirement. The recipient should see or hear the information without searching for it and without using a special tool.

With a graphic the label can sit directly on the image, under it or in the caption shown together with the post. On social media a safe place is the start of the caption, especially when the platform may cut off longer text.

In a video the information can be shown on screen at the start of the film. It should stay visible long enough to be read comfortably. For audio alone, a short spoken message at the start of the recording works.

For artistic, satirical and fictional content the rules let you choose the form of the label so that it does not interfere with how the work is received. The information can go in the description, or in the opening or closing credits, depending on the format. It still has to be noticeable.

Labelling a deepfake does not settle every legal question. The rules on protection of a person’s image, personal rights, personal data and copyright still apply. The label says where the material came from, but it does not give consent to use someone else’s face, voice or creative work.

On my own blog I label illustrations made with AI tools in the image caption. That rule goes beyond the minimum duty for deepfakes, but it makes editorial work simpler and gives the reader clear information.

Chatbots and texts: two simpler rules

A chatbot should tell the user that they are talking to AI, at the latest in the first message. Texts prepared with the help of AI usually do not need a label if a human has checked the content, edited it and takes responsibility for publishing it.

The provider of the chatbot is responsible for making it technically able to pass on that information. A business launching a ready-made product should still check that the message really appears at the right moment.

It can read: “I am an AI assistant and I answer questions about the offer”. For a bot handling orders it is worth spelling out how the user can get in touch with a human.

Information placed only in the site’s terms and conditions comes too late. The user should see it before the conversation starts or in the first message.

The rules on texts are narrower in scope. The labelling duty covers publications generated or altered by AI that inform the public about matters of public interest. These can be materials about politics, public health, safety, consumer rights, finance or events that matter to public debate.

A label is not required when the text has gone through real human review and a named person or business holds editorial responsibility. Checking the spelling is not enough. The editor should assess the meaning, the facts, the sources and the conclusions, and be able to change or reject the material.

Product descriptions, how-to posts and newsletters prepared with AI support can generally be published without a label if they go through that process. The business should be able to name the person who approved the text and is responsible for its final version.

A short internal note with three pieces of information is enough: who edits the content, how fact-checking works and who approves publication. With a bigger team you can add the approval date and the name of the tool used.

What to do this week: a plan for a small business

A basic review will take a small business between about an hour and one afternoon. What usually takes longest is gathering material from several channels and working out which items may meet the definition of a deepfake.

  1. List the places where the business uses AI content: the website, the shop, the blog, social media profiles, the newsletter, adverts, podcasts and video material.
  2. Pick out the images, films and recordings that show real people, products, places or events. Assess whether the material could give the impression of being authentic. Label deepfakes at first contact. For the rest of the content you can adopt one broader labelling rule.
  3. Check the chatbot and any forms that hold an automated conversation. The information about contact with AI should appear before the conversation or in the first message.
  4. Decide who holds editorial responsibility for texts prepared with the help of AI. Write down that person’s full name and the basic scope of the checks before publication.
  5. Prepare a single page of internal rules. Write down which materials you label, how the label reads and where it should appear in each channel.
  6. Separate new publications from the archive. Material generated and shared before 2 August 2026 does not need to be labelled retrospectively. For realistic content that is still being promoted you can add a label voluntarily, starting with adverts and popular posts.

I did a similar review recently for the Polish digital accessibility act and the EU requirements for e-commerce described in the article on online store accessibility. The pattern is the same: an inventory, fixes, a short note.

Fines, supervision and what comes next

Breaching the duties in Article 50 of the AI Act can lead to an administrative fine. The maximum level is EUR 15 million or 3% of the business’s total worldwide turnover for the previous financial year.

For companies, as a rule, the higher of those limits is taken into account. For SMEs, including start-ups, a more favourable ceiling applies: the lower of the amounts or percentage values. The authority also weighs the scale and duration of the breach, its effects, whether it was intentional, the business’s earlier actions and its cooperation during the proceedings.

The Polish ustawa o systemach sztucznej inteligencji (act on artificial intelligence systems) was signed by the president on 24 July 2026 and published on 27 July. It sets up Komisja Rozwoju i Bezpieczeństwa Sztucznej Inteligencji (the Commission for the Development and Security of Artificial Intelligence), or KRiBSI. The provisions on supervision, inspections, proceedings and fines come into force on 28 October 2026.

Article 50 of the EU regulation applies directly from 2 August. The later start date for part of the Polish supervision system does not move the duties of businesses.

Generator providers have their own duty to put machine-readable marks in content. Systems placed on the market before 2 August 2026 were given until 2 December 2026 to adapt. That transition period applies to tool providers and their technical solutions.

A business publishing a deepfake is still responsible for a label that a human can see or hear. The simplest routine for future publications covers a list of the tools in use, one template for the message and a named person who approves content. If you want to apply these requirements to your website or shop and do not know where to start, write to me. I will tell you where labels are needed.

FAQ

Does every AI graphic on a company profile have to be labelled?

The duty in Article 50 applies to deepfakes, that is realistic material showing real people, places or events in a way that could mislead. A fictional illustration generated by AI does not need a label, but one shared rule of labelling all such graphics makes work simpler and builds trust with your audience.

Do blog posts written with the help of AI have to be labelled?

Usually not. If the text has gone through real human editing and a specific person or business holds editorial responsibility for the publication, a label is not required. The duty applies to unedited AI texts that inform the public about matters of public interest.

Does minor photo retouching in an editor with AI features need a label?

No. Adjusting sharpness, white balance or the crop, or removing a small element from the background, is standard editing. Replacing the whole background, pasting in new objects or a change that transforms the meaning of the material goes beyond ordinary editing.

What should labelling AI content look like in practice?

The rules do not impose a wording. Simple labels work: “AI-generated”, “Modified by artificial intelligence”, for audio a short spoken message at the start of the recording, for video information on screen. The label has to be visible on first contact with the material.

What are the fines for missing labels and who enforces them?

A maximum of EUR 15 million or 3% of annual worldwide turnover; for SMEs the lower of those values applies. In Poland supervision is carried out by KRiBSI, set up by the act on artificial intelligence systems, with powers from 28 October 2026.

Does material published before 2 August 2026 have to be labelled retrospectively?

There is no duty to go back through the archive. For realistic material that you are still actively promoting (in adverts, pinned posts, on your website), adding a label voluntarily is a good first step.

Filip Chrapek

Let's work together

What do you need?
Budget
When?

Prefer email? hi@philgood.dev