Must you tell callers they're talking to AI?

Short answer: in the EU, yes from 2 August 2026. Article 50 of the AI Act requires people to be told they are talking to an AI system, unless it is already obvious, with fines up to €15 million or 3% of worldwide turnover. In the US, AI voices already sit under the TCPA’s robocall rules. Details and exact quotes below.

What the EU rule actually says

The duty lives in Article 50(1) of Regulation (EU) 2024/1689, the AI Act, and it lands on the people who build the system. Quoting it exactly: “Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.”

Two things matter in that sentence. First, the design duty sits with the provider (the platform), so the disclosure has to be built in, not bolted on. Second, the escape hatch is the word “obvious”. Here is our read, and it is a position rather than settled law: a phone call is exactly where “obvious” is weakest. The whole selling point of a modern voice agent is that it does not sound like a robot, and the better your agent sounds, the less you can lean on that exception. Relying on “obvious” for a natural-sounding phone agent is a gamble we would not take.

Article 50(5) settles the how and when: the information must come “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure”. On a phone call, that means the top of the call, not a mumble at the end.

When it bites, and what it costs

The transparency chapter becomes applicable on 2 August 2026. You will see a 2 December 2026 date quoted around this topic; read it carefully, because it is narrower than it looks. The Commission’s own consultation notice says that grace applies to generative systems already on the market, for the duty to mark synthetic content. It is not a grace period for telling callers they are talking to AI. If your voice agent takes its first call on 3 August 2026, the disclosure duty applies to it, full stop.

The penalty ceiling comes from Article 99(4), which names “transparency obligations for providers and deployers pursuant to Article 50” in the tier fined “up to EUR 15 000 000 or, if the offender is an undertaking, up to 3% of its total worldwide annual turnover for the preceding financial year, whichever is higher”. Ceilings are not typical outcomes, and national regulators set actual amounts, but the number tells you how seriously the EU takes this specific duty.

The US layer: no disclosure law, but real teeth

The US comes at this from a different angle. There is no single federal “tell the caller it’s AI” statute. What exists is the FCC’s Declaratory Ruling of 8 February 2024, adopted unanimously, which “recognizes calls made with AI-generated voices are ‘artificial’ under the Telephone Consumer Protection Act”. The ruling took effect immediately, and the FCC’s release spells out the consequence: AI-voice calls are “held to those same standards” as robocalls, which for telemarketing means prior express written consent before the call is made.

The enforcement picture is broader than one agency. The TCPA gives the FCC civil fining authority, lets carriers be ordered to block traffic, and, unusually, hands consumers a private right of action, so individuals can sue over unlawful calls. The FCC also flagged that state attorneys general, 26 of whom backed the ruling, can use it directly. So for outbound calling into the US, the practical rule is consent first, and disclosure is simply good sense on top. For inbound answering (a receptionist agent picking up your own phone line), the TCPA is not the issue, and some states add their own bot-disclosure rules, so check the states you operate in.

What your agent should actually say

The compliant version is also the simple version. Put the disclosure in the greeting, in plain words: the business name, then something like an acknowledgement that the caller is speaking with an AI assistant that can help or hand over to a person. That satisfies “clear and distinguishable… at the latest at the time of the first interaction” with no cleverness required. Check where your platform keeps this: most builders let you edit the first message directly, and the platforms we profile with US outbound tooling (A2P registration, TCPA workflows) surface consent handling in their campaign settings. Our vendor pages record each platform’s published compliance posture, and the receptionist roundup and HIPAA roundup flag where the paperwork is real rather than marketing.

Our position, since fence-sitting helps nobody: disclose everywhere, regardless of jurisdiction. The legal duty only exists in some places, but the trust cost of a caller discovering mid-call that the “person” was software is the same everywhere, and a clean disclosure costs you one sentence.

The honest limits

Three things to hold loosely. The Commission’s Article 50 guidelines were still a draft when we captured them (published 8 May 2026, consultation closed 3 June 2026), so the final version may sharpen details like exactly how spoken disclosure should be worded. Member-state enforcement appetites are unknown until regulators start acting after 2 August 2026. And we are a benchmark site, not a law firm: every quote above links to its primary source with a capture date, and the right move for anything high-stakes is showing those sources to your counsel rather than taking our word for it.