EU AI Act Transparency Requirements: Article 13 vs Article 50

Half the posts about EU AI Act transparency circulating right now contradict the other half. One says the transparency rules are delayed to December 2027. Another says transparency becomes mandatory this August. Both are quoting the Act correctly. They are talking about two different Articles with two different requirements that confusingly share a name.

• Article 13: Transparency and Provision of Information to Deployers
• Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems

Requirement	Who you owe it to	Deadline
Article 50 disclosure (chatbot notice, deepfake and biometric disclosure)	The people your AI touches	2 August 2026
Article 50(2) marking of AI-generated content	Anyone consuming the content, machine-readably	2 December 2026
Article 13 transparency and instructions for use (high-risk systems only)	The business deploying your system	2 December 2027

If your AI is customer-facing, the deadline that matters is seven weeks away, not eighteen months. The rest of this piece explains which requirement is which, why the dates split, and what to do about each.

Two requirements, two audiences

Article 50 is what you owe people. If your AI system interacts directly with natural persons, they must be informed they are dealing with AI, “in a clear and distinguishable manner at the latest at the time of the first interaction.” If your system generates synthetic audio, images, video or text, the outputs must be marked as artificially generated in a machine-readable format. Deployers of emotion recognition or biometric categorisation must notify the people exposed, and deployers publishing deepfakes must disclose them. These duties apply whatever the system’s risk classification. A minimal-risk customer service chatbot is fully in scope.

Article 13 is what a provider owes a deployer. It only applies to high-risk AI systems: the Annex III categories like recruitment, credit scoring, and education, plus regulated-product safety components. The system must be designed so the deployer can interpret its output, and it must ship with instructions for use covering capabilities, limitations, accuracy figures, oversight measures, and how to read the logs. The audience is the compliance team at the company running your system, not their customers.

The two articles regulate different relationships. Telling a loan applicant “you are talking to an AI assistant” satisfies part of Article 50. It tells the bank’s risk team nothing about how the scoring model works, which is Article 13’s job.

We’ve already written pieces into both articles: Article 50 and Article 13.

Does this affect you?

Article 50 applies if any of these is true:

1. Your AI system interacts directly with people: a chatbot, a voice agent, an AI email responder.
2. Your system generates synthetic audio, image, video or text content.
3. You deploy emotion recognition or biometric categorisation on people.
4. You publish deepfakes, or AI-generated text on matters of public interest.

Article 13 applies if both of these are true:

1. You are the provider of the AI system (you built it, or you sell it under your name).
2. The system is high-risk under Article 6.

Most businesses with customer-facing AI hit at least one Article 50 trigger. Far fewer carry Article 13 obligations, but those who do carry Article 50 as well, because the obligations are cumulative.

Why the dates split

The original plan was simple: nearly everything applied from 2 August 2026. Then the Digital Omnibus, provisionally agreed on 7 May 2026, moved the Annex III high-risk obligations to 2 December 2027. Article 13 lives inside the high-risk chapter, so it moved with them. Article 50 sits in its own chapter, and it did not move.

The Omnibus then made the marking duty stricter, not looser. The grace period for implementing machine-readable marking of AI-generated content was cut from six months to three, landing the deadline on 2 December 2026.

So when a post says “the AI Act is delayed”, it means the high-risk obligations, including Article 13. When a post says “AI transparency is mandatory from August”, it means Article 50 disclosure. Both are right, and reading either as the whole story is how a compliance plan ends up sixteen months late on the wrong obligation.

One caveat in plain English: the Omnibus is a provisional political agreement awaiting formal adoption and publication in the Official Journal. Until that happens the old dates technically govern, but nobody expects enforcement of the old high-risk date, and the Commission’s draft Article 50 guidelines confirm authorities are preparing to apply Article 50 from 2 August 2026 as scheduled.

When both apply to the same system

Suppose you run a recruitment platform that screens CVs and ranks candidates for employers, with a chatbot that answers candidates’ questions about their application. CV screening is Annex III high-risk. The chatbot talks directly to natural persons.

From 2 August 2026, every candidate who opens that chatbot must be told they are talking to AI, at the start, in the channel they are using. A line in your terms of service fails the test. That obligation is yours now, regardless of the high-risk delay.

By 2 December 2027, the employers using your platform must hold your Article 13 instructions for use: quantified accuracy, known failure modes, the oversight measures you have built in, and how they retrieve the logs they are required to keep. Without that document, your customers cannot meet their own deployer obligations, and the gap traces back to you.

Same product, two transparency obligations, two audiences, deadlines sixteen months apart. The fines sit in the same tier either way: up to €15 million or 3% of global turnover under Article 99, whichever is higher. For a platform with €200 million in revenue, that is a €15 million ceiling per obligation breached, and the cheap one to fix is the one arriving first.

The third transparency

For completeness: people affected by certain high-risk AI decisions also have a right to an explanation under Article 86. That one is owed by the deployer to the individual, and it is covered in the Article 86 piece. If someone on X tells you AI Act transparency means individuals can demand explanations of AI decisions, they are thinking of this, and it is a different obligation again.

What to do now

If your main exposure is a customer-facing chatbot or AI agent: implement the Article 50 disclosure before 2 August 2026. This is days of work, not months, and undisclosed chatbots are the easiest violation a regulator can detect.

If your product generates images, audio, video or text: plan the machine-readable marking work now for 2 December 2026. This is engineering, not copy. Marking that gets stripped by your own pipeline does not comply. The published Code of Practice on Transparency of AI-Generated Content sets out the marking and detection measures providers are expected to meet.

If you provide a high-risk system: you have until 2 December 2027 for Article 13, but the instructions for use depend on your risk management file and testing evidence, which take months to build properly. Use the deep dive to scope it.

If you are not sure which side you are on: classify first. Whether each system is high-risk, and whether you are its provider or deployer, determines which transparency obligations you carry and when. Everything else follows from that.

The confusion online is understandable, because two obligations share a word and a regulation. Your compliance plan does not get to share the confusion. Put the August date on the calendar first.