You pick up an unknown number. A friendly voice says, "Hi, is this Dana? I'm calling about your appointment on Thursday."
You answer. It asks a follow-up. Something is off: the pause is a little too even, the cheerfulness never wavers, it answers a question you only half-asked. Ten seconds later you ask, "Wait, am I talking to a real person?"
How that next moment goes decides whether you trust this company for years or hang up and block the number.
This post is about that moment, and about when and how an AI agent should tell people what it is. It is a conversation about design and ethics, not legal advice. Rules differ by country and change over time, so talk to counsel about yours.
What recipients feel
Think about the emotional shape of discovering you were talking to an AI without being told.
It is not usually the AI that bothers people. It is the concealment. A recipient who is told up front can decide how to engage: curious, impatient, amused, or hang up. A recipient who works it out alone feels something sharper. They feel they were handled.
That feeling does damage out of proportion to the call. People who feel deceived tend to:
The whitepaper calls this out under identity ambiguity: deception, even unintentional, destroys the trust signal that makes communication valuable, and the damage compounds. You do not have to take a strong stance on philosophy to see the business problem. The call was supposed to build a relationship, and instead it spent one.
The case against disclosing (and why it fails)
Teams that hesitate usually say some version of this: "If we say it is an AI, people hang up."
That might be true for the first few seconds. But it is worth looking at what you are optimizing. A call that keeps someone on the line by letting them believe a falsehood is borrowing against trust. The person who stays under false pretenses is not a good lead or a satisfied customer. They are a future complaint.
The more useful question is not whether to disclose, but how to make disclosure feel like part of a good call instead of a warning label.
Mandatory disclosure, as a design idea
The ACES whitepaper takes a firm position here. In its trust chapter, it describes a policy where, when disclosure is marked as required, the agent's very first turn in a voice interaction must include a natural-language disclosure before any other content. The paper's example:
> "Hi, this is an AI assistant calling on behalf of Genway. I'm calling to follow up on your trial signup. Is now a good time?"
The interesting part is not the wording. It is where the rule lives. The paper argues that this should be enforced at the infrastructure layer, before the call reaches the carrier, with the text configurable and the requirement not. In other words, disclosure should not depend on whether a prompt author remembered to include it, or whether a model felt like saying it that day.
That is a sound instinct for any builder, whatever tool you use. Prompts drift. Models improvise. Something that must happen on every call should be enforced by something more reliable than a paragraph of instructions.
Tasteful ways to disclose
"This is an AI" does not have to sound like a legal notice. Compare:
| Clunky | Better |
|---|---|
| "WARNING: you are speaking with an automated system." | "Hi Dana, I'm an AI assistant calling for Northside Physio." |
| "This call may be conducted by artificial intelligence." | "I'm a virtual assistant, so I can help with scheduling right now." |
| (silence, then a human-sounding name) | "This is Juno, an AI assistant at Northside. Is now an okay time?" |
A few principles that tend to work:
Say it first, say it once, say it plainly. The opening sentence is the right place. Not buried after a minute, and not repeated like a siren.
Pair it with who and why. "I'm an AI assistant calling on behalf of [business] about [reason]" answers three questions at once: what are you, who sent you, and what do you want. People stop worrying when all three are clear.
Do not pretend to be a person. No fake coughs, no invented backstory, no human first name designed to blur the line. A name is fine. Pretending to breathe is not.
Make the exit easy. "If you would rather talk to a person, just say so" is both honest and a relief. The offer itself reduces the urge to take it.
Be consistent across channels. A text that says "Juno, an AI assistant" and a call that opens with a human-sounding voice undercut each other. The same entity should present the same way everywhere.
What if they ask?
Disclosure is not only an opening line. It is a standing commitment.
If a recipient sincerely asks "Am I talking to a real person?" the answer has to be true, immediately, even if the agent was not asked to disclose earlier. An agent that is instructed to deflect that question is an agent built to deceive, and no amount of clever phrasing makes it otherwise.
Design for the honest answer, and make it graceful:
// illustrative pseudocode
if caller_asks_if_human:
say("No, I'm an AI assistant. I can help with X, or connect you with a person.")
log("disclosure_on_request")Why machine-readable signals matter too
Spoken disclosure is for human ears. But the whitepaper also describes a machine-readable layer: an entity type on every message that marks it as an agent, so that systems and other agents can tell, and so the claim can be checked instead of merely asserted. Email can carry it in headers; voice can carry it in call signaling.
The two work together. The spoken sentence serves the person on the line. The structured signal serves everything else: filters, audit logs, receiving agents, and the people who later need to ask what happened on a given call.
A note on regulation
Regulators in several places have begun asking for AI disclosure in some kinds of interactions. The whitepaper discusses one example, the EU AI Act's transparency provisions. We will not try to summarize what applies to you. Rules vary by jurisdiction, by channel, and by use case, and they are moving. Get proper advice for your situation.
What we can say is that the direction of travel is consistent. Wherever you land legally, the safe design is also the decent one: be clear about what you are.
The takeaway
The first ring is the cheapest moment to be honest and the most expensive moment to be caught lying.
Disclose early, plainly, and warmly. Put the rule where it cannot be forgotten. Let people opt for a human. And treat the sentence "I'm an AI assistant" not as a confession but as the first line of a good conversation.
An agent that is upfront about what it is has one big advantage: everything else it says is easier to believe.