Do you have to tell customers they’re talking to an AI?

There is a moment, in every company that decides to let software interview its customers, when somebody asks whether the customers need to be told. The question is put delicately. Nobody proposes lying. It is more that “AI” seems an unnecessary word to lead with, like opening a dinner invitation with a list of the allergens.

Behind the delicacy sit two fears. One is legal: is there a rule? The other is commercial: if people know, will they still talk to us, and will they still say what they think?

Both fears are reasonable. Both, as it happens, are out of date.

An experiment in selling

The commercial fear has a birthplace, and it is a call centre. A financial services company once let chatbots telephone more than 6,200 of its customers to sell them a product. The machines were good at it. Kept anonymous, they sold as well as the firm’s experienced staff.

Then the researchers changed one thing. They had the chatbot announce itself at the start of the call. Purchases fell by about 80%. Customers who knew what they were talking to grew curt, hung up early and kept their money.

That number has had a long career. It appears, stripped of its origins, whenever a team debates how candid to be about its AI, and it is always offered as proof that candour is expensive. But consider what was being measured. An uninvited caller was asking people to spend money. Suspicion was the rational response. The study shows that people dislike being sold to by a machine. It shows nothing at all about being listened to by one.

An experiment in listening

Change the purpose of the conversation and the result turns over.

In one large American study, every participant was told from the first message that their interviewer was an AI. Afterwards they were asked whom they would choose next time. Forty-three per cent picked the AI. Nineteen per cent picked a human. The rest didn’t mind. For many of them the machine was the attraction. There was nobody to impress and nobody to disappoint.

This is an old finding in new clothes. In a health screening interview run years earlier, people who believed they were speaking to a computer alone were less afraid to open up than those who thought a person was watching. What makes us guarded is the prospect of being judged. A machine, honestly labelled, cannot judge. We go into this further in whether people tell AI interviewers the truth.

People also want the label, and in large numbers. When Pew asked 5,023 American adults, 76% said it was extremely or very important to be able to tell whether something had been made by AI or by a person. So the customer to worry about is not the one who reads “AI interviewer” and declines. It is the one who works it out afterwards and feels tricked by a company they were trying to help.

A gap remains in the evidence. Nobody has yet run the cleanest test, which would hide the machine from half the participants in a feedback interview and compare what the two halves say. And the people studied so far were mostly paid panel members, a different animal from a busy head of operations. It is worth watching how many customers begin your interview once they have read the invitation. Nothing so far suggests that honesty costs you the conversation.

The law arrives

While researchers were establishing that disclosure does little harm, legislators were deciding that it was compulsory anyway.

Since 2 August 2026, Article 50 of the EU’s AI Act has applied. An AI system built to interact directly with people must be designed so that those people are informed they are dealing with one. The duty falls on the provider, the company that makes the system. If you use somebody else’s interview tool, the obligation to build the notice in is theirs, though it would be an odd customer who didn’t check.

Two details repay attention. There is an exception for cases where it is “obvious” to a reasonably well-informed, observant person that they are talking to an AI. It is a narrower door than it looks. A chat that asks thoughtful follow-up questions in fluent, friendly prose is precisely the situation in which it may not be obvious. And the notice must come “at the latest at the time of the first interaction”. A line in the footer will not do, nor will a confession at the end. The European Commission’s guidelines confirm the August start date.

The AI Act governs the label. An older law, GDPR, governs the answers. The moment a reply can be tied to a person, by an email address, a personal link or a recorded voice, it is personal data, and the familiar duties follow. You need a legal basis, which for customer research usually means consent or legitimate interests. You owe people a privacy notice saying who you are, what you want the data for and how long you will keep it. You need a contract with the vendors who handle it. Voice deserves a word of its own. A recording is personal data, but it becomes the specially protected kind, biometric data, only when it is used to identify the speaker. Transcribing what someone said is one thing. Building a voiceprint is another.

America, characteristically, has no single rule and a growing number of local ones. California’s was written with a particular villain in mind: the bot that conceals its nature in order to deceive someone into a purchase or to sway a vote. A pure feedback interview is neither, so the law does not squarely apply, though an interview that ends in an upgrade offer drifts towards it. Utah’s takes another line. A business using generative AI with consumers must own up whenever a person clearly asks, on pain of fines of up to $2,500 per violation, and an AI that says so plainly from the outset is protected. Maine requires notice wherever a reasonable consumer might think they were talking to a person. Colorado’s law concerns AI used in consequential decisions such as hiring and lending, which a product survey is not. Britain’s Market Research Society tells its members to inform participants when AI affects them.

Voice adds a last complication. Federal law lets a conversation be recorded if one party consents. California wants everyone’s consent. The safe habit is to say so before the first spoken answer and to let people type instead.

None of this is legal advice, and before a large or sensitive study a lawyer should have the final word. But look at what these laws, drafted in different places for different reasons, have in common. Not one of them can be satisfied by keeping quiet.

What to say

So the cautious course and the honest one have converged, which is rarer than it should be. What remains is a matter of manners.

Say it twice: once in the invitation, so that people can decline before they click, and again in the first message. Say what will happen to the answers, who reads them, whether they are tied to the customer’s account, and how long they are kept. If you record voice, say that too. Give a human name, or at least an inbox that a human reads. And if a respondent asks halfway through whether they are talking to a bot, the bot should say yes.

One more thing. Keep the interview free of selling. An upgrade prompt at the end pulls you towards the stricter American rules, and back into that call centre, the one place where we know that honesty hurts.

The message itself can be short.

Change the brackets until every line is true. A promise of anonymity that you cannot keep is worse than no promise.

The delicate question, then, has an indelicate answer. Yes, tell them, and tell them first. The fear that customers will walk away belongs to a study about selling. When the machine is there to listen, the people who know what they are talking to turn out to be the ones most willing to talk.

How Diaform helps

Diaform runs customer interviews as a chat with an AI interviewer that asks your questions one at a time and follows up when an answer is vague. Respondents type or speak, in 32 languages, with no account needed, and every conversation keeps its full transcript beside the summary. What you tell respondents, and your privacy notice, remain yours to write. There is a 14-day free trial.