Voice clones and call-centre AI: the Article 50 disclosure rules in practice
Article 50 makes voice your most visible AI compliance surface. This guide covers IVR disclosure wording, synthetic speech marking, the exemptions and what to log for inspection.
Voice is where EU AI Act compliance becomes audible to your customers. From 2 August 2026, Article 50 requires that anyone reaching your contact centre knows whether they are speaking to a person or a machine, and that synthetic speech is marked in a way software can detect. It is one of the few obligations every mid-market company hits regardless of risk tier: chatbots, IVR menus, outbound calls and cloned voice recordings all fall inside it. This guide covers the call centre AI disclosure requirements in the EU practically — the opening line a voicebot needs, how voice clones must be labelled, where the exemptions genuinely hold, and what to log so an inspection is a five-minute conversation.
Call centre AI disclosure requirements in the EU: Article 50(1)
Article 50(1) is the “you are talking to a machine” rule. It applies to AI systems intended to interact directly with natural persons and requires that the person is informed at the latest at the time of the first interaction or exposure. Three consequences follow:
- The greeting is the disclosure. If the first thing a caller hears is “press 1 for billing” and a synthetic voice appears three menu layers deep, you have disclosed late — the notice belongs at the IVR opening line or the first sentence of an outbound AI call.
- Outbound calls count equally. A voicebot dialling about appointments or arrears must disclose before the conversation it called to have, not in a voicemail disclaimer.
- Silence is not neutrality. No de minimis exception: a five-second automated menu is still an interaction with an AI system.
Two points are routinely missed. The carve-out where the AI nature is obvious from the context is narrower than it reads: a line explicitly labelled automated self-service may qualify; a lifelike voice saying “I’m Sarah from the care team” does not. The honest test: would a reasonable caller be surprised this was a machine? If yes, disclose. And the duty sits on the deployer as much as the provider — buying a voicebot platform does not move it to the vendor.
Synthetic voice disclosure: the Article 50(2) marking rule
Article 50(2) addresses content rather than conversation. Providers generating synthetic audio must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, as far as technically feasible — the rule behind every “ai generated speech marking requirement” in vendor documentation. The work splits in two:
- Providers carry the marking duty. The speech-synthesis platform must embed a detectable signal — C2PA-style content credentials, watermarking or an equivalent. In procurement, ask what marking is applied, in which format, how a third party verifies it.
- Deployers carry the disclosure duty. Using synthetic audio that manipulates what a listener would understand — most obviously a voice clone of a real person — requires disclosing that the content is artificially generated. The AI Act voice clone labelling rules are a public-facing statement wherever that audio is published, not only a label on the file.
Tagging alone is not enough: the manipulation must be disclosed to the listener in an effective way. A silent watermark satisfies the machine half and fails the human half — a cloned figure’s voice in a training recording or investor update needs an audible or captioned statement plus the marker on the file. The compliant path takes an afternoon: confirm the platform marks exports, add a spoken line at the start, record both facts in the campaign file. The obligation follows the deployer, not the invoice.
The exemptions, phrased carefully
Mid-market teams hear “exemption” and stop reading. Three carve-outs exist, and each is narrower than the sales deck suggests:
- Obvious AI (Article 50(1)). Where the AI nature is obvious from context — a line labelled “automated assistant”, a menu naming the self-service route. It does not rescue a realistic human voice that never says what it is.
- Assistive and standard editing (Article 50(2)). Outputs fall outside the deep-fake duty where the AI performs assistive standard editing that does not substantially alter the input — noise reduction, slight pace correction. Re-voicing a script in a cloned timbre fails that test comfortably.
- Artistic, satirical and fictional works. Scripted dramatisations sit outside the deployer duty; a message a customer might take to be genuine account communication does not.
Document the exemption, not the conclusion: one paragraph stating which carve-out applies and why beats a tick-box. Where counsel is unsure, the cost asymmetry decides: a sentence of disclosure costs nothing, a transparency breach does not.
Voicebot legal notice: the EU AI Act opening line that works
Most call centres over-lawyer this and under-perform: a caller will not parse a paragraph, and a notice buried at the end of an IVR tree is not a notice. Three patterns tested against Article 50(1):
- Inbound IVR opening: “You have reached our automated assistant. Say or press for anything, or say ‘agent’ to speak to a person.” First sentence, human route named, works in speech and DTMF alike.
- Outbound AI call: “Hello, this is an automated message about your delivery. Press 1 to continue, or 5 to be removed from this list.” Disclosure precedes the substance; the opt-out is not legally required but is what your contact-centre manager will ask for anyway.
- Handover line: “I’m transferring you to a person from our care team” — and if either leg was synthetically generated, say so there.
Three rules separate a defensible line from a decorative one. Put it first, because the obligation attaches at the first interaction. Make it repeatable: menu loops should re-offer the human route rather than assume the caller heard the opening. Say “automated”, not “AI-powered assistant”: callers understand it, regulators read it as an admission of AI, and the shorter phrase survives a noisy connection.
Do not script evasions: a synthetic voice told to hesitate, or to answer “of course, I’m happy to help!” when asked “am I speaking to a human?”, is the exact pattern the article exists to prevent. Script the honest answer — “I’m an automated assistant; I can connect you to a person if you prefer” — then sample calls weekly to confirm the line is actually played: a compliant script the platform fails to speak is still a breach.
Article 50 voice generator disclosure by scenario: the working table
Most disclosure questions reduce to one of the situations below — the implementation column can be lifted straight into a runbook.
| Scenario | Duty | One-line implementation |
|---|---|---|
| Inbound IVR, voicebot | Disclose AI interaction at first contact (Art. 50(1)) | Prepend “automated assistant” to the greeting; re-offer the human route in menu loops |
| Outbound automated calls | Disclose before the substantive message (Art. 50(1)) | Open with “this is an automated message…”, then purpose and opt-out |
| Lifelike or cloned voice on live calls | Disclose interaction and synthetic manipulation (Art. 50(1), 50(2)) | State the voice is synthetic in the greeting; tag assets with machine-readable marking |
| Cloned brand voice in ads, hold messages | Machine-markable output; manipulation disclosed (Art. 50(2)) | Export with content credentials on; add a spoken or captioned synthetic-voice notice |
| Website chatbot or SMS assistant | Text channels too: inform at first interaction (Art. 50(1)) | Persistent “AI assistant” badge at session start, not only the footer |
| Third-party agency produces the audio | Deployer duties remain yours regardless of who generated it | Contract for machine-readable marking and disclosure evidence in the delivery file |
| Human-recorded greeting | No marking needed; avoid misleading framing | Never script a human greeting to imitate an AI, or an AI to imitate a person |
| Human-edited AI draft | Assistive-editing exemption may apply | Keep the editing record; disclose where the result still misleads |
Delegating production does not delegate disclosure: require the agency to deliver every synthetic asset with its marking intact and a note on where disclosure appeared. Enforcement usually follows not concealment but inconsistency — a disclosure in the terms of service the call flow never plays, an English script’s line missing from the local-language variant, marking on marketing audio but not retained support calls. Pick the disclosure moments, put them in the scripts, audit them like any other control.
Disclosure logs and what a transparency breach costs
When a market-surveillance authority writes, the question is never “do you have a policy?” — it is “show me that this call disclosed on 14 March.” Four log fields survive most inspections:
- What was disclosed — the exact wording that played, versioned; scripts drift, and a log without wording cannot show what a caller heard in February.
- Where and when — channel, line or page, and timestamp.
- Which system — the AI system identifier from your register, so the log cross-references classification, provider and contract.
- The exemption or marking claim, if any — which carve-out or marker applies and who approved it; an unexplained blank turns a file request into an on-site inspection.
Keep it proportionate: the platform stores current disclosure strings with change history, campaigns record a one-line attestation at launch, published audio is checked before release. The first extract typically surfaces campaigns launched with last year’s wording — a morning’s fix that would otherwise surface during an investigation.
On cost: Article 50 sits in the middle penalty tier under Article 99 — €15 million or 3% of global annual turnover, whichever is higher. Starting points vary by Member State; the ceiling does not. Transparency breaches are among the easiest to detect, because the evidence is public. Three factors move a case from theoretical to expensive:
- Systematic absence — every line answered without disclosure looks like a design choice rather than an oversight.
- Concealment — answering “yes, I’m a person” when the system is not converts a transparency gap into misleading conduct, a tier above, inviting wider investigation.
- No remediation record — an authority receiving a corrected script, an updated log and a dated change record on first contact closes matters differently.
Scripts, marking and log together cost an afternoon; the middle tier costs up to 3% of worldwide turnover.
How to check your own position in three minutes
The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.
Frequently asked questions
Do call-centre voicebots have to say they are AI?
Article 50(1) requires people interacting with an AI system to be informed — for a phone channel, an audible disclosure at the start of the interaction, not a line buried in the IVR tree. “Would you like to speak to a human?” does not satisfy it alone; the system’s nature must be stated, not implied by an offer of transfer.
Does cloning a human voice for an advert need labelling?
If the speech is synthetic, Article 50(2) requires the audio to be machine-markable and the manipulation disclosed — an advertisement using a cloned presenter voice falls squarely inside that. Realistic synthetic speech without marking is exactly the harm the provision targets.
What exemptions exist, and how far do they stretch?
The article narrows where the AI’s nature is obvious from context, where the system performs an assistive single task, or where the output is human-edited and disclosed — treat any exemption as something to document, not assume. If a customer could reasonably believe they are hearing a person, you are outside the carve-out regardless of how the exemption reads.
What happens if our AI mistakenly denies a refund?
Transparency duties interact with your redress process: an automated front line generating outcome messages needs a human escalation path, and the automated logic must be recoverable from logs when the complaint arrives. Retain interaction logs and disclosure records together — an inspection will ask when the customer was told, and what the system said afterwards.
How do we audit that the disclosure is actually played?
Treat it as a control, not a script: sample calls weekly across languages and channels, verify the opening disclosure appears in the retained recordings, and check the marking pipeline for outbound synthetic audio. The gap between “the vendor says it discloses” and “the recording proves it” is where transparency breaches live.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
Does the EU AI Act Apply to US Companies? Territorial Scope Explained
10 min
ExplainersEU AI Act for SMEs and Start-Ups: What Applies, What's Exempt, and the Fine Rule That Helps You
9 min
ExplainersThe Digital Omnibus Delay: Your December 2027 High-Risk AI Deadline Is Already Ticking
12 min