Do You Have to Tell Callers They Are Talking to an AI?
Short version: in the EU, yes, and that duty went live on 2 August 2026. In the UK, no blanket rule exists for an inbound business line. In the US there is no federal disclosure law, one state statute genuinely covers voice, and most of what circulates about the rest is misread. This article links every law it names, because the top answers on this question, in forums and in vendor blogs alike, routinely cite statutes that say something else. None of this is legal advice; it is a sourced map to take to your lawyer.
The answer by jurisdiction
| Where | Must an AI on a call announce itself? | The actual instrument |
|---|---|---|
| EU | Yes, unless it is obvious to a reasonable person | AI Act, Article 50(1), applies since 2 August 2026 |
| UK | No blanket duty; transparency rules bite around the edges | ICO guidance, PECR for outbound marketing |
| US federal | No. Outbound AI robocalls need consent, inbound is untouched | FCC ruling 24-17 (2024); the disclosure NPRM is still not final |
| US states | One genuine voice statute, several narrower ones | Maine's chatbot law; Utah on request; California for robocalls |
The EU duty that just switched on
The AI Act's Article 50(1) is direct, and worth quoting rather than paraphrasing: providers must ensure AI systems that interact with people are built so that those people "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." Article 50(5) sets the timing, disclosure "at the latest at the time of the first interaction". The Act entered into force in August 2024, and the transparency chapter became applicable on 2 August 2026, so this stopped being a future obligation a few weeks ago.
Three details that matter on a phone line
Three details matter for a phone agent. The duty sits on the provider, the company that builds and ships the system, though deployers carry adjacent duties. The "obvious from context" escape is a high bar on a phone call, where nothing visual signals AI and the entire product goal is sounding natural. And the reach is extraterritorial by design: Article 2 applies the Act to providers and deployers outside the EU "where the output produced by the AI system is used in the Union", which is exactly what a call with someone in the EU is. The fine tier for transparency violations runs to EUR 15 million or 3% of worldwide turnover, whichever is higher, with a lower-of-the-two softening for SMEs.
The US map, drawn from the statutes
There is no enacted federal law requiring an AI to identify itself on a call. The FCC's February 2024 ruling made AI-generated voices "artificial" under the TCPA, which means outbound robocalls using them need prior express consent, and the ruling's own terms cover systems that initiate outbound calls, not AI answering your inbound line. The FCC's separate proposal to require in-call AI disclosure, from August 2024, remains a proposal; the FCC's own mid-2026 agenda lists its next action as undetermined.
The state layer is where the real rules and the real myths both live:
- Maine enacted the one law that plainly covers AI voice agents in commerce: its chatbot statute defines the technology to include "aural" communication and forbids letting a reasonable consumer believe they are talking to a human without clear notification. No intent to deceive is required, and violations fall under Maine's Unfair Trade Practices Act.
- Utah requires generative AI, expressly including audio, to disclose when a consumer clearly asks, and to disclose proactively in regulated occupations.
- California's phone rule is AB 2905, which requires robocalls placed with automatic dialling to disclose the use of an artificial voice. Outbound, and not an AI-receptionist rule.
- The FTC's Impersonation Rule sits under all of it: an AI agent posing as a bank, an agency or another business is a violation regardless of the technology.
The claims that do not survive the statute text
The folklore is confident and specific, which is what makes it spread. Checked against the primary sources:
| The claim | What the source actually says |
|---|---|
| "California's BOT Act requires AI disclosure on calls" | The Act defines a bot as an automated online account and limits itself to online communication. Phone calls are outside its scope |
| "Federal law requires announcing AI at the start of a call" | No such law exists. The FCC proposal is unfinalised, two years on |
| "Texas requires AI disclosure within 30 seconds" | The 30-second rule is a pre-existing automated-dialler statute. Texas's SB 140 contains no AI provisions at all |
| "Colorado's AI Act requires interaction disclosure" | It never became operative; it was replaced in 2026 before taking effect |
| "The FCC fines AI non-disclosure at $1,500 per call" | That is the TCPA's general wilful-violation figure. There is no AI-disclosure penalty schedule |
New Jersey's bot law, like California's, is online-only. Texas's TRAIGA adds AI disclosure duties for government agencies and healthcare, not general business lines.
The UK position, and what actually bites
No UK statute requires an inbound business AI line to announce itself, and the UK deliberately chose a principles-based approach over an AI Act. What exists instead has edges worth knowing. The ICO's transparency expectations under UK GDPR require telling people how their data is processed, and its automated-decision-making rules, recast as new provisions by the 2025 data reforms, engage where decisions have significant effects, which a table booking does not. For outbound marketing, PECR requires prior consent for automated calling systems, a rule that predates AI and catches it cleanly. Regulated firms answer to the FCA's existing frameworks rather than any new AI rulebook. And the advertising codes cover recorded marketing messages but not live calls, an odd seam worth knowing exists.
This is consistent with what we have published before: the comparison post and the restaurant playbook both state the UK inbound position in one paragraph. This page is the long version.
The second disclosure everyone forgets
Whether the call is used to train the model is a separate question from whether a robot is speaking, and it has sharper teeth in the UK. ICO guidance on purpose limitation requires the training purpose to be specified up front, or a compatibility assessment before recordings are reused, and its generative AI consultation outcomes call for training purposes to be "explicit and specific". The stock phrase "calls may be recorded for quality and training" was written for training staff; reading it to cover training a model is, in our reading of the guidance rather than any explicit ICO ruling, unlikely to survive scrutiny. If recordings feed model improvement, the privacy notice should say so in those words.
What disclosure actually does to callers
Here is where the legal question meets evidence, and where the answer splits by context in a way no compliance article mentions.
Outbound sales, where disclosure costs
For outbound sales, disclosure carries a measurable penalty. The Luo, Tong, Fang and Qu field experiment in Marketing Science, run on real sales calls, found that announcing the bot up front sharply cut purchases, that customers became curt and rated the identical agent as less knowledgeable, and that disclosing after the outcome removed most of the penalty. Nature Machine Intelligence published the mechanism as a transparency-efficiency tradeoff: the same algorithm performs worse purely because people know. The prejudice is the human's, not the machine's.
Inbound service, where disclosure pays
For inbound service, the evidence points the other way. In the vendor Upfirst's analysis of 450,000 AI receptionist calls, agents that disclosed being AI saw roughly 20% fewer hang-ups, and Censuswide's 2026 UK survey for Moneypenny found 68% of consumers comfortable with AI answering once they know a person is reachable. The Journal of Service Management work by Mozafari and colleagues adds the nuance: disclosure damage concentrates in high-stakes moments, and honesty pays most precisely when the bot stumbles.
The upside for your speech recognition
There is also an upside no policy page mentions. Phonetics research by Cohn, Ferenc Segedin and Zellou shows people speak differently to known machines, clearer and more deliberate. A disclosed agent gets easier audio to transcribe. Honesty literally improves your speech recognition.
Hiding it fails on its own terms
Suppose you skipped every law above. Concealment still buys less than it appears to. Peer-reviewed detection studies say listeners cannot reliably spot synthetic speech, 73% accuracy against 2022-era voices in the PLOS ONE study, worse against modern clones in the 2025 Scientific Reports work, so some callers will be fooled. But the same overconfidence cuts both ways: people now accuse humans of being bots, a phenomenon documented across call centres, and the folk tests callers deploy, gibberish, jokes, "what day is it", are unvalidated against current systems. What concealment reliably produces is the gotcha moment, the point mid-call where the caller realises, recalibrates everything the agent said, and leaves feeling tricked. The trust cost lands at the worst time, and under the EU Act it now lands alongside a legal one.
How to word it
The evidence favours early, brief and capability-framed, then moving on:
"Hi, you're through to [business]. I'm their AI assistant, I can book you in, answer questions and take messages. How can I help?"
One sentence, states what it can do rather than apologising for what it is, and satisfies "at the time of the first interaction" wording by construction. Make the greeting uninterruptible if your platform allows it, the settings for that are in the barge-in guide, and keep the recording line separate and honest about training if training is what happens. If you serve EU callers from anywhere, treat Article 50 as applying to you, because by its own words it does.
Frequently asked questions
Does the EU AI Act apply to a UK or US business whose agent talks to EU callers? By the Regulation's text, yes. Article 2 extends it to providers and deployers in third countries where the system's output is used in the Union, and a recital states the anti-circumvention intent plainly. A call answered for or made to someone in the EU is output used in the Union. How enforcement reaches a small non-EU firm is genuinely untested; the obligation on paper is not ambiguous.
Is "obvious from context" a usable exemption on a phone call? It is a high bar. The benchmark person is "reasonably well-informed, observant and circumspect", and the entire commercial point of a modern voice agent is sounding human. A robotic IVR voice announcing menu options is arguably obvious; a natural conversational agent is arguably the opposite. Nothing has been tested before a regulator yet, which is itself a reason not to be the test case.
Do outbound AI marketing calls in the UK need anything extra? Yes, and it is older than AI: PECR requires prior consent for automated calling systems used in direct marketing. That rule, not AI disclosure, is what catches most UK outbound automation, and the ICO enforces it.
What should I do this quarter if I run an AI line? Three things, none requiring a lawyer first. Put the AI identification in the first sentence of the greeting. Check your privacy notice says model training in plain words if recordings are used for it. If any callers are in the EU, confirm your vendor has done its Article 50 work, because the provider duty is theirs and the reputational cost of failure is yours.