Telling people they are talking to a machine

An AI assistant answers customers on WhatsApp and web chat. It has a friendly name, writes fluently, and nowhere says it is software. Customers find out when it misunderstands something obvious — and they feel misled. Since 2 August 2026 this is also a legal problem in the European Union, and comparable rules apply in parts of the United States and elsewhere. The fix is cheap: one honest line, in the right place, backed by a record. This piece covers what the rules ask, how to do it in a messaging thread, and how to make it enforceable in your own systems.

Article 50(1) of the EU AI Act requires that people interacting directly with an AI system are informed that they are dealing with one, unless that is obvious to a reasonably well-informed, observant person. The obligation applies from 2 August 2026. The Digital Omnibus amendments adopted in 2026 delayed other parts of the Act, but legal commentary is consistent that they did not delay this one.

Two details matter more than the headline:

• Timing and clarity. The information must reach the person clearly and distinguishably, at the latest at the first interaction, and meet applicable accessibility requirements. A disclosure buried in terms and conditions, or behind a menu, does not count.
• It stacks. The AI Act does not replace data protection, consumer protection or accessibility law. They apply in parallel; satisfying one does not discharge another.

The EU rule is the broadest, but not the only one.

• California has required bots to disclose themselves since 2019 when they are used to encourage a sale or influence a vote (the B.O.T. Act, SB 1001).
• Utah's Artificial Intelligence Policy Act (2024, narrowed in 2025) requires disclosure in regulated professions and when a consumer asks whether they are talking to AI.
• Other jurisdictions are consulting on or drafting similar rules, and channel policies add their own requirements for automated messaging.

For a programme that operates in several countries, the practical answer is to apply the strictest common rule everywhere: disclose at the start, every time.

A web chat widget can show a permanent label. A messaging thread cannot — there is no frame, only messages. So the disclosure has to be a message, or part of one, at the start.

Do:

1. Open the first automated message with it. Short, plain, not apologetic: "You're chatting with our AI assistant. Type agent at any time to reach a person."
2. Write it in the language of the conversation, in the same voice as everything else.
3. Disclose again after a handback. If a human takes over and later returns the conversation to automation, the person's assumption has changed. Say so again.
4. Make the way out real. "Type agent" must actually reach a person, within stated hours.

Don't:

• Rely on the privacy policy. Nobody reads it, it is not at first interaction, and it is not in the conversation.
• Disclose after the third turn. By then the interaction has happened.
• Let a name imply it. Calling the assistant "Aya" and hoping the customer infers it is not a person is the opposite of clear. A name is fine with the disclosure.

A travel company runs one AI assistant on WhatsApp in English and Arabic, and on its website.

• First message, English: "Hi, I'm Aya, the AI assistant for [brand]. I can help with bookings and changes. Type agent any time to talk to a person (8am–10pm)."
• First message, Arabic: the same content written natively, not translated word for word, with the keyword in Arabic as well as English.
• Handover: "Connecting you to a member of our team now." The human sees the full transcript.
• Handback: "You're back with Aya, our AI assistant." — logged as a second disclosure.
• Record: each disclosure is stored with conversation ID, time, language and exact wording. The send path refuses the first automated reply if the record is missing.

It cost a few lines of copy and a small table. It removed the question entirely.

The most common complaint about automated support is not that it is automated; it is the suspicion of being automated without being told. Disclosure removes that, and what remains is judged on whether it is useful.

It also helps the system work. People who know they are talking to software phrase requests more directly, accept a menu without irritation, and escalate when they need to instead of arguing with something that cannot help. Undisclosed automation gets discovered — usually at the worst moment — and the discovery costs more trust than the disclosure would have.

The instinct is to put the disclosure in the bot's prompt or in the first template. Both work until the second channel arrives, or the third agent, or the vendor changes.

Put it where consent lives. A disclosure is the same kind of fact: a per-person, per-conversation record of something you told them and when.

• Can the disclosure be enforced by the platform — a fixed first message — rather than left to the prompt?
• Is each disclosure logged with time, language and wording, and can I export that log?
• After a handover to a human and back, does the platform disclose again automatically?
• Does the "talk to a person" keyword work in every language and channel I use?

Next in this track: a closer reading of AI Act Article 50 for messaging. For the ledger this record lives in, see consent architecture; for measuring whether the assistant actually helps, read containment is not resolution.