Article 50 of the European regulation on artificial intelligence has applied since 2 August 20261. It does not only concern the companies that build models. It concerns any business that puts an AI system in front of its customers, or publishes what such a system has written.
The text is short. Its consequences for an interface are precise.
Four situations, four obligations
A system that talks with a person must say so. An assistant on the website, an agent answering emails, a voice picking up the phone: the person must know they are dealing with an AI, unless it is obvious from the context. Obvious to you is not the same as obvious to your customer.
Synthetic content must be marked. Providers of systems that generate text, images, audio or video must ensure those outputs are marked in a machine-readable format and detectable as generated. The deadline for that marking is 2 December 20262.
Emotion recognition and biometric categorisation must be announced. Those who deploy them inform the people exposed to them.
A deepfake must be declared, and so must a text of public interest. Content that imitates a real person, place or event must be presented as artificial. So must a generated text published to inform the public on matters of general interest, unless it has been reviewed by a person and someone holds editorial responsibility for it.
What it changes in an interface
The obligation fits in one sentence. Putting it into practice is a matter of detail.
The first line. The notice comes before the first answer, not in the terms and conditions. One plain sentence is enough, in the user's language.
Voice. A telephone agent introduces itself as such in the first few seconds, before asking anything.
Email. A reply drafted by an agent says so in the body of the message, not in a signature nobody reads.
Handing over to a person. When a member of staff takes over the conversation, the interface says so. Transparency works both ways.
The record. Keep proof of what was shown, when, and in which version. That is what you will be asked for.
What you publish
Product pages, newsletters, standard replies, market summaries: a growing share of what a company publishes passes through a model. Two habits are enough to stay on the right side of the text.
The first: mark what is generated, for readers and for machines. A visible notice, and a machine-readable signal in the page.
The second: have it reviewed, and say so. Human review under editorial responsibility is exactly what the text recognises. Our Signals, for instance, carry the notice "Prepared with the help of an AI, reviewed by the house" and a machine-readable mark. None is published without approval.
Outside the Union too
The regulation applies to providers placing a system on the Union market, and to those whose outputs are used in the Union, wherever they are established3. An Israeli company serving French or Spanish customers is as concerned as one based in France.
A short list for this week
List every point where a customer can talk to a system: website, messaging, telephone, email.
Read the first message of each one, in each language. Is it clear, straight away?
Identify what you publish with the help of a model, and who reviews it.
Name one person responsible for these notices, and write down what was decided.
None of this needs a large project. It needs someone to own it, before the question comes from a customer, or from a regulator.
ElahPublished on October 21, 2026