The European Commission published practical guidance on 20 July 2026 for the AI Act’s Article 50 transparency obligations, which begin applying on 2 August 2026. For businesses, the immediate question is not whether every AI use case is high risk. It is whether customer-facing AI, synthetic-content workflows, and public-interest publishing have clear disclosures and defensible operating controls.
What Article 50 puts in scope
Article 50 covers several distinct situations. Providers of systems that interact directly with people generally need to make clear that the person is interacting with AI unless that is obvious in context. Providers of systems that generate synthetic audio, images, video, or text have obligations around machine-readable marking and detectability, subject to technical feasibility. Deployers have separate disclosure duties in specified cases, including deepfakes and certain AI-generated text published to inform the public on matters of public interest.
That means a website chatbot, a content-generation pipeline, and a marketing review process should not be treated as one generic “AI compliance” project. Each has different users, outputs, owners, and evidence needs.
Why the new guidance matters now
The Commission’s guidelines are meant to help providers, deployers, and competent authorities apply Article 50 consistently. The Commission also published a voluntary Code of Practice on transparency of AI-generated content, while making clear that Article 50(2), (4), and (5) apply from 2 August 2026.
For an enterprise deploying AI from third parties, this is a supply-chain task as much as a product task. Teams should record which vendor or internal group is the provider, which business unit is the deployer, what the system produces, and where the required notice or marking is delivered.
A practical pre-deadline checklist
- Inventory interactions: Find public and employee-facing AI experiences, then decide where an AI-interaction notice is necessary rather than assumed.
- Map generated outputs: Separate text, image, audio, and video workflows. Confirm how outputs are marked, what metadata survives downstream tooling, and where technical limits exist.
- Set publication controls: For public-interest content, establish a review path that determines whether disclosure is required and how it will be presented.
- Assign accountability: Name an owner for product notices, an owner for content disclosures, and an owner for vendor evidence.
- Keep implementation evidence: Save configuration decisions, vendor documentation, test records, and change logs. A policy alone does not show how a live AI experience behaves.
What this means for business AI teams
Transparency should become a design requirement, not a footer added at launch. A useful test is simple: can a user understand when they are interacting with AI, and can the business explain how generated content is identified when the obligation applies?
Organizations building chatbots and agents should build the disclosure decision into their launch checklist alongside data access, escalation, and monitoring. That approach makes it easier to update a deployment when models, channels, or regulations change.