EU AI Act Art. 50 Transparency: What Local AI Users Must Do

While the AI Omnibus deal dominated headlines last month, a quieter countdown has been running undisturbed: Article 50 of the EU AI Act — the general transparency obligations — takes effect August 2, 2026. No extension, no transition period. The deadline is fixed.

Adding urgency: a key consultation is closing tomorrow. On May 8, 2026, the European Commission published draft guidelines on implementing the Article 50 transparency obligations, with the consultation period ending June 3, 2026. Law firm Covington & Burling noted on X: "The European Commission has published draft guidelines on the implementation of the transparency obligations under Article 50 of the EU AI Act" — non-binding, but the first authoritative Commission guidance on how Article 50 should be read in practice.

This article explains, based on our reading of the EU AI Act and the Commission's draft guidelines, what Article 50 requires from businesses deploying local AI systems — and where the AI Omnibus delays do and don't apply.

What the AI Omnibus Postponed

The political agreement reached on May 7, 2026 significantly delayed the heaviest compliance requirements. Based on our reading of the published Omnibus agreement:

• Stand-alone high-risk AI (Annex III) — recruitment, credit scoring, education, law enforcement: now December 2, 2027 (pushed from August 2026)
• High-risk AI embedded in regulated products (Annex I) — medical devices, machinery, vehicles: August 2, 2028
• Watermarking obligation for AI-generated media content (Art. 50(2)): postponed to December 2, 2026

Rémy Schlich summarised on X: "EU lawmakers have agreed to delay key obligations for high-risk AI systems."

These postponements, however, apply only to high-risk systems and the media watermarking provision. Article 50 paragraphs 1, 3, and 4 are unaffected by the Omnibus and come into force on schedule on August 2, 2026.

What Article 50 Requires from August 2, 2026

Article 50, based on our reading, establishes four transparency obligations — three of which take effect in August:

Para. 1 — Chatbot Disclosure (from August 2, 2026)

Anyone deploying an AI system that directly interacts with natural persons — a chatbot, a virtual assistant, an automated customer support agent — must ensure those persons are informed at the start of the interaction that they are communicating with an AI system, not a human.

Exception: when it is obvious from context, the obligation falls away. A clearly robotic interface, or an internal tool where all users know they are operating an AI tool, generally falls outside the scope of this obligation based on our reading.

For customer-facing chatbots — on a company website, in live chat, or in a customer portal — disclosure is mandatory.

Para. 2 — Watermarking (postponed to December 2, 2026)

Providers of AI systems that generate synthetic images, audio, or video must technically mark that content. Not relevant for pure language models without media generation capability — the typical SMB LLM stack.

Para. 3 — Emotion Recognition and Biometric Categorisation (from August 2, 2026)

Systems that detect emotions or categorise persons by identity characteristics — voice tone analysis, facial expression processing, behavioural classifiers — must inform exposed persons. Not relevant for standard LLM installations without speech or image analysis components.

Para. 4 — Deepfake Labelling (from August 2, 2026)

AI-generated content that deceptively depicts real persons or events must be labelled as AI-generated, unless it is obviously art or satire. Not relevant for text-only applications without media generation.

Which Use Cases Trigger Art. 50(1) in Practice

The key criterion, based on our reading: does the system directly interact with external persons who cannot reasonably expect to be talking to an AI?

Practical example: A logistics SMB in the Netherlands runs a local LLM assistant on its intranet — employees use it to query internal processes and shipping documentation. → No disclosure obligation: the AI nature is evident to all users. If the same company later deploys a customer-facing chatbot on its website to handle delivery enquiries, that changes: → Disclosure is mandatory.

What the Commission's Draft Guidelines Clarify

The May 8, 2026 draft guidelines, based on our reading, confirm:

• The disclosure obligation sits with the deployer (the business running the system), not only with the model or software provider
• Disclosure must happen at the start of each interaction — not hidden in terms of service, not in a privacy policy, not on a separate "about" page
• The phrasing must be clear and understandable — "You are chatting with an AI assistant" is more effective than vague disclaimers like "AI-powered"
• Language: ideally the language of the user in question

Final guidelines are expected in June 2026, ahead of the August deadline. With the consultation closing tomorrow, a July 2026 adoption is realistic.

Practical Implementation Steps for SMBs

Based on our reading of Article 50 and the draft guidelines, the following steps are prudent before August 2, 2026:

1. Map your AI touchpoints Which systems in your business communicate directly with external users — customers, prospects, applicants, patients?

2. Add disclosure messages For external chatbots and assistants, configure a visible opening message:

"Welcome! You are chatting with [Company Name]'s AI assistant. How can I help you today?"

3. Check for emotion or biometric analysis Is any speech or facial sentiment analysis in use? If so, plan a disclosure routine for exposed persons.

4. Document your classification Record in writing why each application is or is not subject to the disclosure obligation. This simplifies future compliance checks and internal audits.

5. Get legal advice on borderline cases Applications where it is unclear whether external users can reasonably be expected to know about the AI nature should be reviewed by a qualified legal professional.

Why Local AI Simplifies Compliance

Businesses running their AI assistant on their own hardware — via Ollama, Open WebUI, or a comparable stack — have full control over the conversation interface. The disclosure can be added in minutes to the system prompt or the welcome message:

You are [Company]'s AI assistant.
At the start of every conversation, inform the user they are interacting
with an AI system, not a human representative.

With cloud-based solutions, the ability to customise disclosure behaviour depends on what the provider allows — configuration access, API restrictions, and contract terms can delay implementation. Local AI operators face no such dependency.

More on how local AI structurally simplifies regulatory control: Data sovereignty with local AI. The kAIra toolkit includes ready-to-deploy disclosure templates for Open WebUI configurations. For a broader overview of how a local AI pilot can be structured for compliance from day one: Freshlab Pilot Project.

All Current Deadlines at a Glance

The AI Omnibus substantially eased the high-risk compliance burden — and that is welcome relief for SMBs that might have been caught by recruitment or credit-scoring AI classifications. But Article 50 is not delayed. For businesses with customer-facing AI interfaces, the technical implementation is manageable. Two months remain. The question is whether your deployment is ready.

Questions about classifying your AI applications or implementing the disclosure? Get in touch.