SMEs must label AI content if they allow people to interact with a chatbot, use deceptively realistic images, videos or audio recordings, or publish unchecked AI-generated texts on matters of public interest. These obligations apply from 2 August 2026. However, a text that has been reviewed by a human with relevant expertise does not need to be labelled if the company assumes editorial responsibility.
When must SMEs label AI content?
The type of content, its intended use and the company’s responsibility are what matter. The rules do not apply across the board to every piece of content created with the involvement of artificial intelligence.
Article 50 of the EU AI Act applies from 2 August 2026. According to its announcement on the transparency obligations, the European Commission published the final guidelines on practical implementation on 20 July 2026.
A company is a deployer if it uses an AI system professionally and under its own responsibility. According to the European Commission’s guidance, employees who use ChatGPT, Microsoft Copilot or another tool as part of their work are generally not considered deployers in their own right. Responsibility usually remains with the company.
Under the European Commission’s Implementing Regulation, content generated before 2 August 2026 does not need to be labelled retrospectively. Only providers of existing generative AI systems have a transitional period until 2 December 2026 to implement technical, machine-readable labelling.
Chatbots must disclose their artificial identity early
People must be able to recognise that they are communicating with an AI system no later than the start of their first interaction. A separate notice is only unnecessary if this would be obvious to a reasonably attentive person.
This applies, for example, to a support chat on a website, a digital appointment assistant or an avatar in a customer portal. A clear notice could read: You are chatting with an AI-assisted service. It should appear before or together with the first response, rather than only in the legal notice or privacy information.
Under Article 50, the obligation initially falls on the provider of the interactive AI system. If a business develops a chatbot itself or has it offered under its own name, it may itself be the provider. A company integrating a ready-made service should nevertheless check that the required notice is actually displayed and remains visible on mobile devices.
When introducing such a system, it is useful to combine the notice with clear responsibilities, data protection measures and a handover to a human. The page on AI and the automation of recurring work shows how these applications can be incorporated into existing processes.
Realistic images, videos and voices need a notice
AI-generated or manipulated images, audio recordings and videos must be labelled if they could credibly depict real people, places, objects or events. The information must be visible or audible, as appropriate, and understandable to the person concerned.
One example is an advertising video in which an artificially generated person speaks about a product as if they were a real customer. An altered voice recording of the managing director or a realistic image of an incident that never occurred may also fall within the rules. A visible notice such as Image generated using AI, or a corresponding announcement at the start of an audio recording, provides clarity.
Not every AI-generated image is automatically a deepfake that must be labelled. An obviously fantastical illustration that does not appear to be a recording of a real person or event will generally not fall under this obligation for deployers. However, providers of generative systems must also mark such content in a machine-readable format so that technical systems can identify its origin.
Special requirements concerning the form of disclosure apply to art, satire and comparable works. The notice must not unduly impair the presentation or use of the work. Whether a particular approach is sufficient therefore depends on the individual case.
Public AI-generated texts do not always need to be labelled
Labelling is required only if an AI system generates or alters texts that inform the public about matters of public interest. According to the guidelines, these include health, safety, the environment, politics, the economy and consumer affairs.
An automatically published guide on health risks, funding, energy prices or consumer rights may therefore need to be labelled. A short product description, an internal email or a draft quotation is not covered solely because ChatGPT was involved. The specific content and intended publication remain decisive.
Labelling is not required if a human with relevant expertise reviews the text and an individual or the company assumes editorial responsibility. This means that someone should check the statements, figures, sources, conclusions and wording, and amend the text if necessary. According to the European Commission, a spelling or grammar check alone is not sufficient.
A short approval process is therefore advisable for day-to-day operations. Record who reviews public texts, which sources were used and who approves publication. This will also help if questions are raised about a statement at a later date.
Companies should review their use of AI in five steps now
An initial review does not need to begin with an extensive set of rules. A table listing the tool, purpose, person responsible, place of publication and required labelling is sufficient for an initial overview.
- Record applications: Include website chats, text tools, image generators, voice services, video applications and AI functions within existing software.
- Determine your role: Establish whether your company is a provider or deployer, and which service provider supplies the technical labelling.
- Categorise content: Distinguish between chatbots, realistic media, public information texts and purely internal applications.
- Define notices: Specify the wording, placement and timing of each notice. Check mobile views and audio tracks as well.
- Document approvals: Appoint a person with relevant expertise to take responsibility for public texts, and avoid automatic publication without review.
According to the European Commission, infringements may result in fines of up to 15 million euros or 3 per cent of worldwide annual turnover. The exact amount depends on factors including the nature, severity and duration of the infringement, as well as the upper limits that apply to small and medium-sized enterprises. According to the Bundesregierung (German Federal Government) information on implementing the AI Act, the Bundesnetzagentur (Federal Network Agency) will play a central role in coordination and market surveillance in Germany.
ibx company advises small and medium-sized enterprises in Hannover and the surrounding region on how to address such requirements in practice. If you would like to review your applications and processes systematically, you can arrange a no-obligation initial consultation. This provides an opportunity to establish where action is actually needed, without any pressure to make a decision.