EU AI Act labelling rules from August 2026: what companies must do for chatbots, deepfakes and AI-generated text
On August 2, 2026, key transparency obligations under the EU AI Act become applicable. Any organisation operating a chatbot, publishing synthetic media or using AI-generated text for matters of public interest now needs to determine where a visible disclosure is needed, where machine-readable marking is required and who owns each step.
For most SMEs, this does not need to become a months-long legal programme. A reliable AI inventory, clear roles, standard labels and a logged approval workflow can turn an abstract requirement into a manageable process. This guide translates Article 50 into practical implementation without treating every use of AI as if it needed a warning label.
Executive summary
- Chatbots and voice assistants: people need to know they are interacting with AI unless that is already obvious.
- Generative systems: providers need to make relevant outputs technically or machine-readably detectable as artificially generated or manipulated.
- Deepfakes: professional deployers generally need to disclose their artificial nature visibly.
- Public-interest text: AI-generated or materially manipulated text published to inform the public may require visible disclosure, subject to the Act's exceptions.
- Deadline: Article 50 applies from August 2, 2026. The new EU Code of Practice is voluntary but provides a structured implementation framework.
What actually changes on August 2, 2026
The AI Act separates transparency for people from detectability for technical systems. That distinction matters: a visible note beneath an image and machine-readable provenance data solve different problems, and an organisation may need both depending on its role.
Visible disclosure
A person receives a clear and timely notice, such as “You are chatting with an AI assistant” or “This image was artificially generated”.
Technical marking
Metadata, watermarks or other robust techniques make artificially generated or manipulated content machine-readable and detectable.
Not every AI-assisted task automatically becomes a labelling case. An internal draft, a spelling correction or a routine summary is not the same as a deepfake or an AI-generated public-interest article. The key factors are role, content, publication purpose and the degree of AI alteration.
Provider or deployer: your role matters
Many implementation mistakes happen because organisations look only at the tool. The AI Act first looks at the organisation's role in a specific use case.
You are a deployer
You use someone else's system under your authority, such as an AI chatbot on your website or an image model used by marketing. This is the most common role for SMEs.
Practical impact: user notices, visible labels, approvals and correct use of the provider's marking features sit with your organisation.
You are a provider
You develop an AI system, have it developed or place it on the market under your own name. A white-label assistant or custom SaaS product with generative functionality may put you in this role.
Practical impact: technical marking, documentation and transparency features need to be designed into the product.
A note on custom AI agents
A company can be both provider and deployer. If you offer an agent to customers under your own brand, do not infer your role solely from the underlying foundation model. Role assessment belongs in the architecture phase alongside privacy, permissions and secure data access through MCP.
A practical labelling matrix
| Use case | What to plan for | Practical pattern |
|---|---|---|
| Website chatbot | AI interaction notice | “Lyron AI Assistant” at the chat entry point |
| AI phone assistant | Disclosure at the beginning of the call | Short spoken notice before the first question |
| Synthetic video of a real person | Visible deepfake disclosure; assess technical marking | Notice in the video and its description |
| AI text on a matter of public interest | Assess visible disclosure | Editorial note attached to the publication |
| Internal email draft | Usually no public label; follow internal policy | Human review before sending |
| Custom generative SaaS feature | Assess provider duties and machine-readable marking | Provenance metadata and a documented output pipeline |
This matrix is an orientation tool, not a blanket legal assessment. Legal exceptions and adapted forms of disclosure may apply to journalistic, artistic, satirical or editorially controlled content. Document not only that you apply a label, but also why a use case was classified that way.
The 30-day SME checklist
Time is short. This order prevents teams from polishing disclosure copy before they even know which systems are in scope.
Week 1
Map AI systems and roles
- List every internal and external AI system
- Group outputs by text, image, audio, video and interaction
- Record provider, deployer or dual role for each use case
- Assign an owner in the business and in IT
Week 2
Define labels and user experience
- Create standard wording for chat, phone and content
- Place notices where the interaction or publication starts
- Add editorial review for borderline cases
- Check accessibility and all required languages
Week 3
Test technology and suppliers
- Test provider marking and export features
- Preserve metadata through storage, transformation and delivery
- Review contracts and documentation for responsibilities
- Define a fallback if technical marking is stripped
Week 4
Log, test and train
- Run test cases for every relevant channel
- Log the decision, model version, approval and publication time
- Train marketing, service, HR and product teams by role
- Schedule a quarterly review for model and process changes
Build transparency into the workflow
Disclosure cannot depend on individual memory. A robust publication workflow uses four controls:
- 1. Classify: identify channel, media type, intended purpose and owning team.
- 2. Mark: add required metadata and visible notices from a controlled template.
- 3. Approve: route sensitive cases to a person while standard cases follow clear rules.
- 4. Evidence: log model, prompt version, processing, approval and final output.
With n8n, Power Automate or a custom application, this control path can sit inside existing CMS, support and approval workflows. The tool is secondary; transparency, access and evidence need to be designed together.
Minimum audit record
System and model version · responsible person · purpose and audience · type of disclosure · approval status · publication time · link or hash of the final output.
Frequently asked questions
Does every text created with ChatGPT need an AI label?
No. Visible disclosure for text mainly concerns AI-generated or materially manipulated publications intended to inform the public about matters of public interest. Internal drafts and ordinary business communications do not automatically fall into that category. Providers of generative systems have separate technical marking duties.
Does an AI chatbot on a company website need disclosure?
People generally need to be informed when they interact with an AI system unless this is already obvious. A clear “AI assistant” label at the point of entry is timely, understandable and user-friendly.
Do the rules apply to SMEs?
Company size alone does not create a general exemption. The relevant factors are the role, system and use case. A small company can be a deployer, provider or both.
Is the EU Code of Practice mandatory?
No. Signing is voluntary, but Article 50 still applies. Organisations that do not use the Code should document their alternative measures and their effectiveness particularly clearly.
When do the rules apply?
Article 50 applies from August 2, 2026. Proposed transitional measures may still change. Existing systems should therefore be prepared for the regular deadline now.
Official sources and article status
This article was last reviewed on July 15, 2026. The Regulation and current European Commission guidance remain authoritative:
- Regulation (EU) 2024/1689 – EU AI Act
- European Commission: Code of Practice for marking and labelling AI-generated content
- FAQ on signing and implementing the Code of Practice
- European Commission: Navigating the AI Act
Note: This guide provides practical orientation and does not constitute legal advice. Obtain qualified legal review for sensitive or unclear use cases.
Make your AI workflow traceable before August
We identify where your AI systems interact with people or publish content, then build disclosure, approval and audit logs directly into the process.
Book a free consultation