AI Chatbot Disclosure Requirements Under the EU AI Act (Article 50(1))

From 2 August 2026, if your organisation runs a chatbot, voice assistant or any AI that talks to people in the European Union, one rule applies before almost anything else: the system must make clear it is an AI, not a human. That is Article 50(1) of the EU AI Act — the interactive-AI disclosure obligation — and it is the transparency duty most businesses will meet first, because conversational AI is everywhere.

This guide covers exactly who is bound, what the disclosure must look like, the narrow “obvious from the circumstances” exception that companies routinely overestimate, and how to comply without guesswork.

What Article 50(1) requires

Article 50(1) of Regulation (EU) 2024/1689 requires providers of AI systems intended to interact directly with people to design those systems so that the people concerned are informed they are interacting with an AI — unless that is obvious from the circumstances and context of use. The duty sits on the provider: the organisation that develops the conversational system and places it on the market.

The rationale is consumer protection. People behave differently with a machine than with a human — what they disclose, how they interpret answers, how much they trust a recommendation. The law gives them the basic fact they need to calibrate: am I talking to a person or not?

What counts as an “interactive AI system”

The draft guidelines read this broadly. Article 50(1) covers conversational agents and voice assistants in any context — customer support, e-commerce, finance, public services, healthcare, education — as well as AI companions and avatars, humanoid robots and cobots, agentic AI and coding agents, bots on social networks, and AI-driven non-player characters in video games where the interaction is not obvious.

One important clarification for the new wave of agentic AI: where a provider cannot reliably predict whether an agent will end up interacting with a human, the guidelines say it should be designed to disclose its AI nature in every such situation. You do not get to assume no human is on the other end.

The “obvious” exception — narrower than you think

The single exception is that disclosure is not required where the AI nature of the interaction is obvious. But the guidelines define “obvious” against a specific legal benchmark, and it is stricter than the everyday meaning of the word.

The test: obvious from the point of view of a natural person who is “reasonably well-informed, observant and circumspect” — the “average consumer” standard borrowed from EU consumer-protection law. The guidelines then apply a multi-factor assessment:

• The target audience. Who actually uses the system?
• Vulnerable groups. Whether children, elderly people, or persons with disabilities may be part of that audience — their presence raises the bar.
• AI and digital literacy. The level of AI familiarity among the intended users.

The guidelines give concrete examples. A code-assistance chatbot available only to professional developers, and AI-enabled NPCs inside a video game, may meet the obviousness threshold — the audience plainly understands it is AI. A general customer-service or e-commerce chatbot facing the public almost certainly does not: a reasonable member of the public may genuinely be unsure whether they are talking to a person, so disclosure is required.

The practical advice across the board: do not over-rely on this exception. When in doubt, disclose. The cost of a one-line “you’re chatting with an AI” notice is trivial; the cost of wrongly assuming obviousness is a breach.

How to disclose correctly

The regulation does not dictate a single format, but the guidelines are clear on the qualities a disclosure must have — and on what fails.

It must be:

• Made before or at the start of the interaction — for a chatbot, before or at the very beginning of the conversation. Not surfaced later, not left to be discovered.
• Clear, distinguishable and in plain language, and conforming to applicable accessibility requirements.
• Not buried in terms of service, a privacy policy, or documentation. A disclosure the user has to go looking for is not a disclosure.

A brief, visible notice such as “You are chatting with an AI assistant” at the opening of the conversation is the baseline. For voice, an equivalent audible statement at the start of the call.

Sensitive contexts need more. The guidelines flag that where users may express emotional distress or vulnerability, or are likely to form an emotional attachment (think AI companions, mental-health-adjacent tools, services aimed at children), a single first-turn statement is unlikely to be enough. Build periodic, context-aware reminders into the experience so the AI nature stays salient.

Provider vs. deployer — and why you might be both

Article 50(1) is a provider obligation — it binds whoever develops and places the system on the market, because disclosure has to be designed into the system. But many businesses are also deployers: if you operate a chatbot on a customer’s behalf, or deliver a conversational service directly to end users, deployer obligations under Article 50(3) (emotion recognition / biometric categorisation) and 50(4) (deepfakes / public-interest text) can attach to you separately. Map your role per system, not per company.

It reaches non-EU businesses too

Article 50 applies wherever the output is used in the EU. A UK or US company running a customer-service chatbot that serves EU customers engages Article 50(1) just as an EU company does. Establishment outside the EU is not a shield.

Penalties

Breaching the Article 50 transparency obligations falls in the AI Act’s middle penalty tier: up to €15 million or 3% of total worldwide annual turnover, whichever is higher. The figures make the point that disclosure is not an optional nicety.

A compliance checklist for chatbot disclosure

• Inventory every conversational AI you provide or operate — web chat, in-app assistants, voice/IVR, social bots, agents.
• Decide the “obvious” question honestly for each, against the average-consumer standard and your real audience (including vulnerable users). Default to disclosing.
• Implement an opening disclosure — visible text or audible statement, before or at the first turn, in plain language.
• Add periodic reminders in sensitive or attachment-prone contexts.
• Keep it accessible — meet applicable accessibility requirements so the disclosure reaches all users.
• Maintain a disclosure register — for each system, record the mechanism, when it was implemented, who approved it, and the rationale for any “obvious” exception claimed. This register is your evidence if a national authority asks.

Frequently asked questions

Does a clearly labelled chatbot still need a disclosure?

If the AI nature is genuinely obvious to a reasonably well-informed, observant and circumspect member of your actual audience, the exception can apply — but “we labelled it on the product page” is weaker than a disclosure at the start of the conversation. The safe path is an opening notice regardless.

Who is responsible — us or the chatbot vendor?

The 50(1) duty is on the provider that develops and places the system on the market. If you build it, it is you; if you buy it, the vendor carries the design duty, but you may hold deployer duties and should ensure the disclosure is actually switched on in your deployment. Clarify this in your contract.

Do internal-only AI tools count?

The obligation centres on interaction with natural persons; a narrowly scoped internal tool used only by informed professionals may meet the obviousness threshold. Construe this narrowly — if the tool can reach a broader or external audience, disclose.

Is a single “I’m an AI” message at the start enough?

Often yes, but not always. In sensitive contexts — emotional distress, likely attachment, audiences including children — the guidelines indicate a one-time statement may not suffice; build periodic reminders.

When does this apply?

2 August 2026 under the law in force as of June 2026. The Digital Omnibus reform leaves Article 50 on this schedule; confirm the current position before relying on the date.

The bottom line

Article 50(1) is the simplest transparency duty to satisfy and the easiest to get wrong by over-trusting the “obvious” exception. If people might reasonably wonder whether they are talking to a human, tell them they are not — clearly, at the start, in plain language. It costs almost nothing to implement and removes one of the most visible compliance risks your AI products carry.