Does My Chatbot Have to Say It’s AI?

Usually yes, and as soon as possible. The notice has to arrive with the first exchange at the latest, which rules out disclosing it only after someone has typed. The exemption for when it is “obvious” is quite narrow.

Support bots answer the first-line questions at most companies now. Whether the bot has to announce itself usually gets decided by whoever installed the widget, which means it usually gets decided by nobody.

What Article 50 actually says

Article 50(1) of the EU AI Act requires that AI systems intended to interact directly with people are designed and developed so that those people are informed they are interacting with an AI system. There is one carve-out: unless that is obvious to a person who is reasonably well-informed, observant and circumspect, taking the circumstances and context of use into account.

The design duty sits with the provider

Article 50(1) is addressed to whoever builds the system, not to whoever installs it. In practice that distinction protects nobody. The complaint arrives at the company whose website the bot is on. And if you rename the bot, give it a persona, retrain it on your own content or put your brand on it, you may have stepped into the provider role yourself.

If you bought the bot, the practical question is whether the vendor built the notice in and whether your configuration switched it off.

What does “Obvious” mean?

A widget labelled Automated help is close to obvious. A bot called Ava, with a photo, a first-person voice and a typing indicator, is the opposite. The test is not whether a developer would know. It is whether an ordinary attentive person would.

When does the notice have to appear

Article 50(5) settles the timing. The information has to be given in a clear and distinguishable manner at the latest at the time of the first interaction or exposure.

Where that bites is a bot that discloses in its reply, after the person has already typed a question. By then the interaction has happened. A bot that opens the conversation itself with the notice is on firmer ground, because the notice arrives with the first exchange rather than after it.

The safest build removes the argument entirely. Put it in the window, visible before anything happens, and the timing question never comes up.

Where to put it

• In the chat window itself, visible in the opening state before the first message.
• In plain words, not behind a link to terms and conditions.
• Visible on a phone without scrolling. Most chat widgets are opened on mobile.
• Kept accurate at handover. If a human takes over mid-conversation, say that too. The person is entitled to know which one they are talking to at any point.

When you probably do not need it

• The interface makes it unmistakable. A search box is not a chatbot. A form is not a chatbot.
• Law enforcement systems authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to safeguards. This exemption falls away if the system is available for the public to report an offence.
• Internal tools that never reach anyone outside the company. Article 50 is about what reaches other people.

Does the platform’s own AI badge count

Several chat platforms attach their own assistant label. That is useful and it is not a substitute.

You do not control whether the badge renders, it can disappear when the widget is restyled or embedded differently, and it is applied by a third party on their own terms. Put your own line in the window as well.

What to write

Short, factual, in the window before the conversation starts:

Keep it: in the window, not after the first question, and not only in the terms. The same timing logic applies to AI images.

What still catches people

Showing the notice is the easy half. Showing that you showed it, to a specific person, on a specific date, a year after the conversation, is the part nobody has solved.

It gets sharper if the bot does more than answer questions. A bot that screens, scores or routes someone is making a decision about them, which brings Article 22 of the GDPR into it as well: the right to human intervention, to express a view, and to contest the outcome. The same notice ends up serving two regimes, and the record behind it has to serve both too.

That is what Art50 records: which notice a person saw, when, on which system, and what the AI decided afterwards. Sealed with a qualified eIDAS timestamp, so the date and the content cannot be edited later. The free plan covers twenty records a month, no card.

Art50 provides evidence, not legal advice. See the EU AI Act timeline and the full text of Article 50.