"Do we have to tell users they're talking to AI?" is the most common EU AI Act question B2B SaaS teams ask, and the answer is usually yes. Article 50(1) of Regulation (EU) 2024/1689 requires providers to design AI systems intended to interact directly with natural persons so those persons are informed they are interacting with AI — unless that is obvious in context. The duty becomes applicable on 2 August 2026.

This guide is the chatbot-specific deep dive of our Article 50 transparency guide. It covers who carries the duty, how to reason about the obviousness exception, when the disclosure must appear, adaptable wording for four common surfaces, and how to evidence compliance.

What Article 50(1) actually requires

Article 50(1) is a design duty: providers must ensure that AI systems intended to interact directly with natural persons are designed and developed so that those persons are informed they are interacting with an AI system. The exception: no disclosure is required where this is obvious from the point of view of a natural person who is reasonably observant, taking into account the circumstances and the context of use.

Three points follow from that wording. First, the duty attaches to systems intended to interact directly with people — customer-support chatbots, voice agents, conversational copilots — not to backend systems with no direct human interaction. Second, it is about the system's design, not a one-off notice: the disclosure has to be part of how the product presents itself. Third, the exception is context-dependent, which is why it deserves its own section below.

Authoritative text: Regulation (EU) 2024/1689 on EUR-Lex.

Who the duty falls on: provider, not deployer

Article 50(1) sits with the provider — the party that develops the AI system or has it developed and offers it under its own name or trademark. If you ship a chatbot as part of your SaaS product, you are likely the provider of that system for Article 50(1) purposes, even where the underlying model is a third-party API. The deployer-side duties in Article 50 — informing people exposed to emotion-recognition or biometric-categorisation systems under 50(3), and disclosing deepfakes and certain public-interest text under 50(4) — are different obligations covered in the Article 50 pillar guide.

Two practical cautions. Companies outside the EU are not automatically out of scope: the Act can reach providers placing AI systems on the EU market and certain operators outside the EU where the system's output is used in the EU, so scope depends on your role and deployment rather than your registered address. And a SaaS company can hold different roles for different systems — record the analysis per system. Our EU AI Act compliance checklist walks through the role analysis step by step.

The obviousness exception, and how to reason about it

The exception asks whether the AI nature of the interaction is obvious to a reasonably observant person in the circumstances and context of use. That is a fact-specific test, not a checkbox. Two contrasting examples show the shape of it:

Likely obvious: a branded, clearly labelled AI assistant — the widget says "AI Assistant", the product markets the feature as AI, and the user deliberately opens it to talk to the AI. Here the context itself communicates the AI nature.
Not obvious: an AI agent answering a general support inbox or chat channel that customers reasonably assume is staffed by humans. Nothing in the context signals AI, so the exception is hard to sustain.

Most real products sit between those poles, and whether the exception applies is a judgment call. The defensible approach: decide, document the reasoning (what the user sees, what expectations the channel creates, why a reasonably observant person would or would not recognise the AI), and confirm the position with counsel. In practice many teams disclose anyway — a short notice costs little, while relying on an undocumented exception is the position that is hardest to defend in a buyer review or regulator inquiry.

Disclosure wording examples for four surfaces

The Act does not prescribe a form of words. The examples below are adaptable starting points, not legal language — have counsel and an accessibility reviewer adapt them to your product, channels, and supported languages.

Chat widget: a persistent header or banner such as "You're chatting with [Product]'s AI assistant", visible before the user sends a first message, with an option to reach a person where one exists.

Voice agent: an opening line such as "You've reached [Company]. I'm an automated AI assistant — I can transfer you to a person at any point." Spoken disclosure matters here because there is no screen to label.

Email assistant: a footer such as "This reply was drafted by [Company]'s AI support assistant. Reply to reach our team directly." If AI answers the inbox autonomously, disclose in the first reply, not only in a signature page.

In-product AI feature: a first-open notice such as "Responses in this panel are generated by an AI system" alongside a persistent "AI" marker on the feature itself.

The Article 50 pillar guide has companion wording for synthetic-media labels and biometric notices; keep the style consistent across surfaces so users learn to recognise your disclosures.

When the disclosure must appear

Article 50(5) requires the information to be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. For a chatbot, that means the person should know they are talking to AI before or as the conversation starts — a label visible when the widget opens, an opening line on a voice call, a notice in the first automated email. Disclosure that appears only in terms of service, only after several exchanges, or only on request does not meet the standard the text describes.

"Clear and distinguishable" also does work: the notice should be noticeable in the interface where the interaction happens, in language the user population can understand, and accessible to users of assistive technology.

Adjacent Article 50 duties in brief

The chatbot disclosure is one of four Article 50 duties, and teams that trigger 50(1) often trigger others:

50(2) — machine-readable marking (provider). If your product generates synthetic audio, image, video, or text, outputs must be marked in a machine-readable, detectable format, with a carve-out for assistive standard editing.
50(3) — biometric and emotion-recognition notices (deployer). People exposed to emotion-recognition or biometric-categorisation systems must be informed.
50(4) — deepfake and public-interest text disclosure (deployer). Qualifying deepfakes and certain AI-generated public-interest text require disclosure.

Each has its own scope questions and exceptions — the full breakdown, including the voluntary Code of Practice and Commission guidance, is in the Article 50 transparency guide.

Penalties in brief

Non-compliance with Article 50 can fall under Article 99(4), which provides maximum administrative fines of EUR 15 million or 3% of total worldwide annual turnover, subject to the Regulation's rules for undertakings and SMEs; actual penalties are determined case by case by national market surveillance authorities. The full three-tier structure is in our guide to EU AI Act fines and penalties. For most B2B SaaS teams the nearer-term pressure is commercial: a missing chatbot disclosure is one of the few Article 50 gaps a prospect can spot from the outside.

Evidencing compliance: make the disclosure auditable

A disclosure that exists but cannot be evidenced still loses buyer reviews. Four records turn the implementation into something you can show:

Screenshot the disclosure on each surface (widget, voice script, email footer, in-product notice), dated and versioned, and refresh the capture when the UI changes.
Record where and when it appears — which surfaces carry it, at what point in the interaction, in which languages, and how it is exposed to assistive technology.
Log the design decision — which systems you assessed against Article 50(1), your provider/deployer role per system, and who approved the implementation.
Document any exception reasoning. If you rely on obviousness anywhere, record the rationale and the counsel review, and revisit it when the product or channel changes.

Keeping this with your AI inventory and change history means the next security questionnaire starts from a record rather than a scramble.

First step: confirm your role

Whether Article 50(1) is your duty depends on whether you act as provider or deployer for each system. The free assessment gives you a suggested classification and risk-exposure result in about three minutes — not a legal determination.

Run the free assessment

AI chatbot disclosure FAQ

Do I have to tell users they are talking to AI?

Usually, yes. Article 50(1) of the EU AI Act requires providers to design AI systems intended to interact directly with natural persons so those persons are informed they are interacting with AI — unless that is obvious from the point of view of a reasonably observant natural person, taking into account the circumstances and context of use. The duties apply from 2 August 2026. Whether the exception covers your system is a judgment call to confirm with counsel.

We build our chatbot on a third-party model API. Is the disclosure duty ours?

It can be. Offering a chatbot under your own name or trademark generally points toward a provider role for that system, even when the underlying model is a third-party API — and Article 50(1) is a provider design duty. Record a role analysis per system rather than assuming the model vendor’s obligations cover your product.

Is a visible "AI assistant" label enough to rely on the obviousness exception?

It may support the argument that the AI nature is obvious, but the exception depends on the circumstances and context of use, not the label alone. A defensible position combines clear labelling, an upfront notice at first interaction, and a recorded rationale — relying on the exception without documenting why is the weaker posture.

When exactly must the disclosure appear?

Article 50(5) requires the information to be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. In practice: before or at the moment the person starts interacting with the system, not buried in terms or shown after the conversation ends.

What are the penalties for missing a chatbot disclosure?

Keep the disclosure record in one place

Govarna keeps your AI inventory, role analysis, disclosure evidence, and change history together — so "show us your Article 50 analysis" has an answer.

Start your free trial