The European Commission posted it on X last week: "From Aug 2026, EU law will require clear labelling on: deepfakes, AI-generated content, chatbots". That deadline is now less than four weeks away.
Article 50 of the EU AI Act (Regulation EU 2024/1689) becomes applicable on 2 August 2026, exactly 24 months after the Regulation entered into force on 1 August 2024. For any organisation operating an AI system that interacts with users, generates public-facing content, or produces synthetic media, the window to implement compliance measures is closing fast.
What Article 50 Actually Requires
Article 50 establishes four categories of transparency obligation for providers and deployers of AI systems:
Art. 50(1), Chatbot and AI Interaction Disclosure
Providers of AI systems designed to interact directly with natural persons must ensure those systems are configured so that users are informed before or at the very beginning of the interaction that they are talking to an AI system. The only exception is when it is obvious from the context, which in practice covers almost nothing: a live-chat widget on a business website is never obviously an AI without an explicit label.
The obligation rests with both providers (who develop or place the system on the market) and deployers (who operate it in their own professional context). Most SMBs are deployers.
Implementations that satisfy the requirement, based on our reading of the Regulation and the AI Office's published guidance:
- A visible "AI Assistant" or "Powered by AI" badge on the chat widget
- An automated opening message: "Hello, I am an AI assistant. How can I help you?"
- A persistent label in the interface header throughout the session
Art. 50(2), Deepfake Labelling
AI systems that generate synthetic images, audio, or video depicting real people, or realistic-seeming scenes, must label that content as AI-generated. This covers AI-generated product photography, voice synthesis in customer service, and marketing video produced with generative tools.
Art. 50(3), Machine-Readable Marking of AI-Generated Text
When AI systems generate text made available to the public for informational purposes, news articles, product descriptions, press releases, blog posts, that text must carry a machine-readable mark indicating its AI origin. C2PA-compatible metadata is the leading practical standard for this purpose, with the AI Office having published a Code of Practice on marking and labelling.
Art. 50(4), Realistic AI-Generated Audio and Video
AI-generated video and audio that does not immediately appear synthetic must be labelled both visibly (for the human viewer or listener) and in file metadata (for automated systems and platforms).
Provider vs. Deployer: Who Must Do What
The Regulation draws a clear line between two roles:
Providers are entities that develop an AI system and place it on the market or put it into service. If you built a proprietary model and licence it to others, you are a provider. Providers must design systems so that the transparency mechanism exists and can be enabled.
Deployers are entities that operate an AI system under their own authority in a professional context. If you install open-weight models via Ollama and connect them to your business's customer-facing chat interface, you are a deployer. Deployers must ensure that the required disclosure actually reaches the end user.
If the provider's system does not include the required disclosure mechanism, the responsibility shifts to the deployer to implement it at the interface layer. There is no gap in the chain: someone must be accountable, and the Regulation is designed to ensure it is the party closest to the end user.
The Local LLM Misconception
A common assumption among SMBs running local LLM stacks: on-premise deployment provides an exemption from the EU AI Act, because no data leaves your own infrastructure.
Based on our reading of the Regulation, this is incorrect.
Article 50 conditions its obligations on the nature of the user-facing interaction, not on where inference takes place. If your Ollama instance serving Llama 3.3 on a Mac Studio in your own server room powers a customer-facing chat widget, you are a deployer subject to Article 50(1), regardless of whether the model runs in a Frankfurt data centre or your basement.
The good news for local AI operators: compliance is structurally simpler. You own the interface and can modify it today. There is no SaaS vendor to wait on, no support ticket to file, no contractual negotiation required. Adding a label to an Open WebUI deployment takes hours, not weeks.
Practical Checklist for SMBs
Based on our reading of Article 50 and the guidance published by the EU AI Office, here are the steps to complete before 2 August 2026:
Step 1: Inventory your AI-facing systems<br> List every tool where an AI system communicates directly with users: chat widgets, email autoresponders generating AI replies, phone bots, internal helpdesks, HR assistants, document Q&A tools. All fall under Art. 50(1).
Step 2: Implement disclosure at each touchpoint<br> For each system: add a visible label, configure an automated first message, or both. The disclosure must be present before or at the start of every user session. A disclosure buried in a footer or in terms of service does not satisfy the requirement.
Step 3: Review your content generation workflows<br> Any AI-generated text, image, video, or audio your organisation publishes externally may fall under Articles 50(2), 50(3), or 50(4). Walk through your content pipeline and apply the appropriate marking at each output stage.
Step 4: Document your measures<br> Article 50 does not explicitly require documentation of compliance steps, but maintaining a record, which systems were reviewed, what changes were made, and when, is prudent. National supervisory authorities are becoming active in enforcement, and a paper trail is your first line of defence.
Step 5: Brief your team<br> Anyone who publishes AI-generated content externally needs to understand the new labelling requirements before they go live.
Enforcement Exposure
Violations of the EU AI Act's transparency obligations can attract fines of up to EUR 15 million or 3% of global annual turnover, whichever is higher, based on our reading of the Regulation's sanction framework. National supervisory authorities are expected to apply proportionate standards when dealing with SMBs, but the obligation is real and enforcement mechanisms are operational from 2 August.
Why Local AI Architecture Simplifies Compliance
Organisations running local AI infrastructure have full control of their implementation stack. No SaaS vendor update cycle. No contractual negotiation. No waiting for a cloud provider to roll out compliant UI changes on their own timeline.
Companies using cloud-based AI tooling often have no direct route to modify the chat interface: they depend on their SaaS vendor to implement the required labels, and that update may or may not arrive before August 2. Local operators face no such dependency.
For organisations currently evaluating a move from cloud-based AI tools to a locally hosted stack, Article 50 compliance is one more concrete reason: you own the interface, so you control the disclosure.
Freshlab designs and deploys local AI systems built for regulatory compliance from day one. If you need a rapid audit of your current AI systems against Article 50 requirements, or help implementing the required changes, contact us before the 2 August deadline.