Limited-Risk AI Under the EU AI Act: Article 50 Transparency Obligations

If your AI system interacts with people, generates synthetic content, or runs emotion recognition, the EU AI Act imposes disclosure duties on you — even if your system is nowhere near Annex III. This is the limited-risk tier: no conformity assessment, no risk management system, no Annex IV documentation pack. Just transparency. The obligations sit in Article 50, they apply from 2 August 2026, and they were not deferred by the Digital Omnibus.

What Limited Risk Actually Means

The EU AI Act (Regulation (EU) 2024/1689) sorts AI systems into four tiers. Unacceptable-risk practices are banned under Article 5. High-risk systems face the full compliance stack under Articles 6, 9–15, 43, and 49. Minimal-risk systems — most AI — carry no mandatory obligations at all.

Limited risk sits between high risk and minimal risk. The category exists because certain AI deployments create a specific, targeted harm: people interact with a machine without knowing it, or receive synthetic content without being told. Article 50 is the fix. It does not require you to prove your system is safe before launch (that is conformity assessment under Article 43, which only applies to high-risk). It requires you to disclose.

Do not confuse limited-risk transparency with Article 13, which governs transparency obligations owed by high-risk AI providers to deployers — a different relationship and a different set of duties entirely. Article 50 governs disclosures to the people who encounter or use the system.

The Four Duties Under Article 50

Article 50 contains four distinct obligations, each targeting a specific actor and situation.

Article 50(1): Chatbot and conversational AI disclosure

Providers of AI systems designed to interact directly with natural persons must ensure those persons are informed they are interacting with an AI. The disclosure must happen before or at the latest at the start of the interaction. The only exception: when it is obvious from the context that the person is interacting with an AI.

This covers customer service chatbots, virtual assistants, AI-generated phone agents, and any interface where an AI converses with a person. If your SaaS product includes a chat feature powered by a third-party model, you are the deployer and the obligation still applies to you under Article 50(1) — the duty runs to the person on the receiving end regardless of who built the underlying model.

Article 50(2): Machine-readable marking of synthetic content

Providers of AI systems that generate synthetic audio, image, video, or text content must mark that output in a machine-readable format so it can be detected as artificially generated or manipulated. The technical standard for what counts as adequate machine-readable marking is being developed; expect guidance from the AI Office. This obligation targets the AI system provider, not the person who distributes the content downstream.

The scope is intentionally broad. A stock-image generator, a voice-cloning tool, a text-synthesis API — all fall here. Purely functional outputs (predictive text completion, OCR correction) are out of scope; the focus is on content that could plausibly be mistaken for genuine human-created material.

Article 50(3): Emotion recognition and biometric categorisation

Deployers of AI systems that perform emotion recognition or biometric categorisation must inform the natural persons exposed to those systems, in advance. Two exceptions: scientific research, and strictly necessary security purposes such as crime detection.

Note the role shift: this duty falls on the deployer, not the provider. If you are a retailer running a camera-based emotion-detection system at checkout, you must inform your customers. The emotion-recognition example is particularly important because emotion recognition in workplaces and educational settings is prohibited under Article 5(1)(f) — Article 50(3) only applies in contexts where emotion recognition is lawful at all.

Article 50(4): Deepfakes and AI-generated public-interest content

Deployers who use AI to generate or manipulate image, audio, or video content that constitutes a deepfake — realistic-seeming content depicting real persons saying or doing things they did not say or do — must disclose that the content is artificially generated or manipulated. The same duty applies to deployers generating AI-produced text on matters of public interest (news, politics, policy) published for public dissemination.

Exceptions exist for clearly artistic, satirical, or fictional works — but the exception only applies if the person or publication makes the AI-generated nature evident in another appropriate way.

What Article 50 Does Not Require

Article 50 is disclosure-only. It does not require:

• A risk management system (Article 9 — high-risk only)
• Technical documentation under Annex IV (Article 11 — high-risk only)
• Conformity assessment before placing the system on the market (Article 43 — high-risk only)
• Registration in the EU database (Article 49 — high-risk and certain GPAI)
• A Fundamental Rights Impact Assessment (Article 27 — limited to high-risk deployers in certain contexts)

This matters in practice. A company running a customer support chatbot owes the people it talks to a disclosure that they are speaking with an AI. It does not owe a notified body a third-party evaluation. The compliance burden is real but proportionate.

Who Is Covered and When

Article 50 applies from 2 August 2026 — the date of general application under Article 113. Unlike the high-risk Annex III obligations (deferred to 2 December 2027 for stand-alone systems under the Digital Omnibus), Article 50 was not deferred. If your system is in scope, August 2026 is the hard date.

The obligations apply to providers and deployers established in the EU, and to providers and deployers established outside the EU whose systems are used by people in the EU — the Act's territorial scope follows the GDPR model.

Providers are companies that develop an AI system and place it on the market or put it into service under their own name. If you build a chatbot and license it to clients, Article 50(1) machine-readable marking is primarily your obligation.

Deployers are companies that use an AI system in a professional context under their authority. The emotion-recognition and deepfake disclosure duties under Article 50(3) and (4) fall primarily on deployers — the people running the system toward actual users.

GPAI Models Are Not Limited-Risk

General-purpose AI models (GPAI) — large language models and foundation models — are governed by Chapter V of the Act (Articles 51–56), which has applied since 2 August 2025. GPAI is a separate, cross-cutting category. A GPAI model provider has obligations under Article 53 (technical documentation, downstream information, copyright policy, training-data summary) regardless of what risk tier any downstream application falls into. Article 50 applies to specific deployments that use a model to interact with people, not to the model itself. These are different layers.

Penalties for Breach

Failure to comply with Article 50 is a breach of "other obligations" under Article 99(4): the fine ceiling is €15 million or 3% of total worldwide annual turnover, whichever is higher. Under Article 99(6), SMEs and start-ups benefit from a proportionality rule — the fine is capped at the lower of the fixed amount or the percentage.

That ceiling is the same as the one for most high-risk provider and deployer failures. Limited risk does not mean low stakes.

Practical Examples

Customer support chatbot. A 60-person SaaS company runs a chat widget on its product. The widget is powered by a third-party model but presented as the company's assistant. Article 50(1) requires that users know they are talking to an AI before the conversation begins. A short disclosure line at the top of the chat window — "You are chatting with an AI assistant" — satisfies this. No conformity assessment. No technical documentation. One sentence.

AI image generator. A marketing agency uses a generative image tool to create campaign visuals and distributes them publicly. The tool provider must ensure those images carry a machine-readable marker under Article 50(2). The agency deploying them in a political or public-interest context also owes a visible disclosure under Article 50(4) if the images depict real people.

Emotion-sensing retail analytics. A supermarket chain deploys a camera system that classifies customer emotions at checkout to measure satisfaction. Article 50(3) requires the chain (as deployer) to inform customers in advance — signage, a notice at entry. The chain should also verify the system does not stray into the prohibited territory of Article 5(1)(f), which bans emotion recognition in workplaces; here the context is customers, not workers, so Article 50 applies rather than a prohibition.

How Confir Helps

Article 50 has a short compliance checklist, but the first question — does Article 50 apply to this system at all? — requires walking the four-part trigger structure systematically. Confir's classification engine uses a rule-based, deterministic intake to answer exactly that: it evaluates whether your system interacts with natural persons (50(1)), generates synthetic content (50(2)), uses emotion recognition or biometric categorisation (50(3)), or generates deepfakes or public-interest text (50(4)), and flags which duty fires.

For companies managing a portfolio of AI tools, Confir's AI register surfaces which systems carry Article 50 obligations alongside any high-risk duties, so nothing is missed at the August 2026 deadline.

Frequently Asked Questions

Does Article 50 apply to my internal-only AI tools?

Article 50(1), (3), and (4) apply when an AI system interacts with, or affects, natural persons — including employees. If you deploy emotion recognition or a deepfake tool that targets your own workforce, the duties apply. The "obvious from context" exception in Article 50(1) is narrow; it does not cover all internal deployments automatically.

We use a third-party chatbot provider. Who bears the Article 50(1) duty?

Both can be in scope. The provider of the underlying model bears the obligation to design disclosure into the system. The deployer (your company) bears the obligation to ensure the disclosure actually reaches the user. If your vendor's chatbot does not disclose its AI nature and you deploy it to your customers, you carry the compliance risk.

Is a verbal or written disclaimer at sign-up enough for Article 50(1)?

Probably not on its own. The obligation requires informing the person at the time of interaction. A buried clause in terms and conditions signed months earlier is unlikely to satisfy the "before or at the start of the interaction" requirement. The disclosure must be timely and visible.

Can I rely on Article 50's artistic exception to avoid the deepfake disclosure?

The exception in Article 50(4) applies when the AI-generated or manipulated nature "is evident from the context" of the work. Clearly satirical cartoons or acknowledged fictional films qualify. Marketing materials, news-adjacent content, or any output where the AI origin might be mistaken for reality do not qualify for the exception. If you are not certain, disclose.

What is the difference between Article 50 limited-risk and Article 13 high-risk transparency?

Article 13 requires high-risk AI providers to give deployers information about the system — its capabilities, limitations, intended purpose, and instructions for use. It is a business-to-business disclosure, part of the high-risk compliance stack. Article 50 requires providers and deployers to inform end users — the people actually interacting with or affected by the system. The two articles address different relationships and different obligations.

Related guides

• Article 6 risk classification pathways
• OpenAI API compliance assessment
• Article 6 high-risk definitions
• Article 6 classification decision tree
• Article 8 compliance requirements
• compliance checklist for Articles 6–29