AI Disclosure Requirements in the EU: Full Scope
Most discussion of EU AI disclosure requirements focuses on chatbots, because that's the most common case — but Article 50's disclosure duty actually covers a wider set of AI system types, and knowing the full scope matters if your product touches any of them. This page covers which systems trigger a disclosure requirement, what specifically has to be disclosed, and the edge cases teams usually miss.
Which systems trigger disclosure
Chatbots and virtual assistants
Any conversational interface — support widgets, sales bots, in-app assistants, voice agents — that interacts directly with a person.
Emotion recognition systems
Systems that infer or classify a person's emotional state from their behavior, expressions, or voice trigger a disclosure duty to the people being analyzed.
Biometric categorization systems
Systems that categorize people based on biometric data (excluding standard biometric identification/verification) carry their own disclosure obligation to the people being categorized.
AI-generated content shown to the public
Content generation itself doesn't require disclosure to the generating user in the same way, but the published output has separate labeling obligations, covered on our content labeling page.
What “disclosure” specifically means
The requirement is that a person is told they're interacting with an AI system, or being analyzed by one, unless that's already obvious from context. It's a disclosure duty, not a consent requirement — the person doesn't need to agree or opt in, they need to be informed. That distinction matters for implementation: a disclosure notice with an “I understand” button the user must click satisfies the letter of the requirement, but so does a clear, persistent notice with no interaction required at all.
The “obvious from context” exception
Disclosure isn't required when AI involvement is obvious from context — a model playground, a demo explicitly framed as “try our AI,” or a research tool where the only thing on offer is a model. This exception is narrower than most teams assume: a support widget with a human name and a friendly, human-sounding tone doesn't qualify just because the team building it always knew it was a language model. The exception is about what's obvious to the person on the receiving end, not to the team that built it.
Disclosure requirements don't stand alone
Disclosure is one of three obligations under Article 50, not the whole of it. Content that's AI-generated has its own labeling requirement, and being able to prove disclosure actually happened is a separate audit-trail requirement. A product can satisfy disclosure perfectly and still be non-compliant if it can't produce evidence that the disclosure was shown, or if generated content reaching the public isn't labeled.
How Disclarion covers the full scope
Disclarion's SDK logs disclosure events per interaction with the jurisdiction and interaction type attached, so coverage can be checked against whichever obligation actually applies — including surfacing gaps where a system type is being used without a matching disclosure rule configured, across OpenAI, Anthropic, and Gemini.
Frequently asked questions
Who has to receive the disclosure — the end user, or a regulator?
The end user, at the point of interaction. A regulator may later ask for evidence that disclosure happened, which is a separate audit-trail requirement, not the disclosure itself.
Does disclosure need to happen every time, or just once?
The Act specifies the obligation, not a frequency. In practice, a clear notice at the start of an interaction, followed by a persistent but non-intrusive indicator, satisfies the intent without becoming ignorable wallpaper.
Are B2B AI tools exempt from disclosure requirements?
Not automatically. If a B2B tool's output or interaction ultimately reaches a person — including an employee of the business customer — the disclosure logic still applies to that interaction.
What language does the disclosure need to be in?
The Act doesn't specify this directly, but the general principle of a disclosure actually being understood implies it should be in a language the user can reasonably be expected to read.