Voice Cloning Scams: How They Work and How to Stay Safe

Voice cloning scams use AI-generated audio of someone you know, such as a relative or a boss, to pressure you into sending money or information. The FBI warns that a clone can sound nearly identical to the real person, so your ear is not a reliable check. Call back on a number you already trust, and agree a secret word with family in advance.

How do voice cloning scams work?

Most of these scams follow the same four steps.

Collect audio. The scammer needs a sample of the voice they want to copy. Vendors and researchers publish very small sample sizes, as the next section shows.
Clone the voice. Software turns the sample into a model that can say new sentences in that voice.
Make a pressure call. The caller invents an urgent problem. The FTC calls these fake emergencies, and it says they are built to stop you from thinking clearly.
Ask for hard-to-trace payment. The FTC lists wire transfers, cryptocurrency, payment apps and gift card numbers as the routes scammers ask for.

How much audio does a voice clone need?

The figures below come from banks, researchers and vendors. They describe what a system can do. They are not tests of how well a fake would fool a particular listener.

| Source | Claim | Type | Date | |---|---|---|---| | Starling Bank campaign | A voice can be replicated from "as little as three seconds of audio" | Bank campaign page | 2024-09-20 | | Microsoft VALL-E paper (arXiv 2301.02111) | Personalised speech from "a 3-second enrolled recording" | Authors' research claim | Submitted 2023-01-05 | | Qwen3-TTS README | "3-second rapid voice clone" | Open-source project README | Released 2026-01-22 | | Google Gemini 3.8 TTS | Voice replication uses "a 30-second audio sample" plus a verbal consent recording | Vendor product rules | 2026-09-23 |

The last row shows what a responsible vendor asks for. Google says users must provide "a verbal consent recording from the voice owner" before a voice can be replicated. The lesson for families is simpler than the numbers. According to these claims, a short recording of someone's voice can be enough to make a convincing copy, so a voice on a phone call is not proof of identity.

Real cases

Three documented cases show the range, from corporate fraud to an election robocall and an impersonation of a senior official.

| Case | When | What was faked | Outcome in reports | |---|---|---|---| | Arup, Hong Kong office | Reported May 2024 | A video call with AI replicas of colleagues, including the chief financial officer. This was video and voice, not voice alone. | About HK$200 million (about US$25.6 million) sent in 15 transactions. Arup told CNN that "fake voices and images were used". Reports from 2024 described the investigation as ongoing. | | New Hampshire primary robocall | January 2024 | An AI voice of President Biden telling voters not to go to the polls | The FCC announced a $6 million fine against political operative Steve Kramer on 2024-09-26 | | Impersonation of Marco Rubio | Mid-June 2025 | A Signal account with the display name marco.rubio@state.gov, which was not a working address, contacted at least five senior officials and sent AI voice messages to at least two | Known from press accounts of a State Department cable. No one is named in the reports reviewed. |

How common are these scams?

The FTC publishes the best official count, but it does not measure voice cloning. Its June 2026 release on 2025 data reports these figures:

Total reported fraud losses of $15.9 billion from 3 million reports.
Imposter scams were the most-reported fraud category, at nearly one in three fraud reports, with $3.5 billion in reported losses.
Business impersonators caused nearly $1 billion in losses, up from $866 million in 2024.
Government impersonators caused about $920 million in losses, up from $789 million in 2024.

The FTC pages do not mention AI or voice cloning. Use these numbers as context for impersonation in general, not as a count of voice scams.

A Starling Bank survey published in September 2024 found that 28% of UK adults believed they had been targeted by an AI voice-cloning scam in the past year. Eight percent said they would send whatever was asked, even if the call seemed strange. The survey was commissioned by the bank, and the page does not show its sample size.

Why you cannot rely on your ear

Two studies point the same way. In a PLOS One paper from September 2025, 150 participants listened to human voices and to voice clones made with ElevenLabs. Clones were labelled human 58% of the time against 62% for real voices in one experiment, and 70% against 72% in another. The gaps were small, and generic AI voices were mostly spotted.

A May 2026 preprint by Mueller and Choong, which has not been peer-reviewed as far as we could check, found a different problem. Accuracy on real recordings fell from 72.7% in a 2021 baseline to 64.1%. Listeners became more suspicious of genuine speech, and commercial systems were the hardest to spot, at 61.3% to 65.9% accuracy.

The practical lesson is that an accent, a pause or a familiar phrase is not proof. The FBI says a clone and the real contact can sound nearly identical.

How to protect yourself and your family

These steps come from the FBI and the FTC. None of them depends on spotting a fake voice.

Hang up and call back. Find the person's number yourself, from your contacts or an old bill, and call that number. Do not use the number the caller gives you.
Agree a secret word or phrase. The FBI advises families to create one to verify each other's identity. Starling advises never sharing it digitally.
Ask a question only the real person would know. The FTC suggests this if you cannot hang up.
Call another family member or a friend. The FTC advises doing this even if the caller asks you to keep the call secret.
Never pay by wire, gift card or cryptocurrency to someone you know only by phone or online. The FBI gives this advice directly.
Turn on multi-factor authentication and never share a one-time code with anyone who calls you.

No source we checked tests how well a secret word holds up under pressure. Treat it as one layer of protection, not a guarantee.

What businesses should change

Voice alone should not approve money or grant access. In July 2025, Sam Altman told a Federal Reserve conference that some financial institutions still accept a voiceprint as authentication, and that AI "has fully defeated that". That was one executive's view, not a measured test, but it matches the FBI advice above. Common controls for a company include:

A callback to a number already on file before any payment is approved.
A second approver for transfers, especially to new payees or for urgent requests.
No single voice or voice-only login as the sole approval for an account.

What the rules say

Regulators are moving, but slowly, and the rules differ by country.

Robocalls in the US. In February 2024 the FCC ruled that AI-generated voices count as "artificial" under the Telephone Consumer Protection Act. Robocalls that use them therefore need the called party's prior express consent, unless an exemption applies. The FCC described this as clarifying existing rules.
Impersonation rules. In September 2026 the FTC voted 2-0 to seek public comment on whether to update its Impersonation Rule to address online platforms' advertising tools. That notice does not address AI voices or impersonation of individuals.
Tool safeguards. In March 2025, Consumer Reports assessed six voice-cloning products and found that four did not take basic steps against unauthorised cloning. It recommended a technical way to confirm a speaker's consent.

This is general information, not legal advice.

Using voice tools responsibly

If you make a custom voice, use a recording you have permission to use. OndaLabs' custom voices are made that way: from a recording you have permission to use, or designed from a text description. Commercial use is included on Starter and above.

For more on the legal side, read voice cloning laws and what voice cloning is. To check whether a clip is synthetic, read AI voice detection.

Frequently asked questions

How can I tell if a call is an AI voice clone?

Listening is not enough. The FBI says a clone can sound nearly identical to the real person. Hang up, call back on a number you already know, and ask a question only the real person would know.

Should my family have a safe word?

Official sources recommend one. The FBI advises creating a secret word or phrase with family members to verify identity. No source we checked tests how reliably it works, so use it alongside a callback.

What should I do if a relative calls asking for money?

Do not send money during the call. Hang up, contact the relative on a number you know is right, and call another family member or a friend. The FTC advises against keeping the call secret, even when the caller asks you to.

Can AI fool a bank's voice login?

Some executives have said so. In July 2025 Sam Altman said AI "has fully defeated" voiceprint authentication, a view he offered as an opinion rather than a measured test. Banks should not rely on voice alone, and neither should you.

Are AI voice robocalls illegal in the US?

Not always. The FCC ruled in February 2024 that AI voices count as artificial under the TCPA, so these robocalls need prior consent unless an exemption applies. Check the current rules for your situation, and ask a lawyer about a specific case.