How do you tell if a voice recording or call is AI cloned?

Why the scam works

The US Federal Trade Commission's consumer alert on family emergency scams describes the pattern: a panicked call that sounds exactly like a relative, a crash or an arrest, and a demand for money. The FTC notes that a scammer needs only a short audio clip, which can come from content posted online, plus a voice cloning program. Its advice: do not trust the voice. Call the person on a number you know is theirs, or reach them through another family member. Requests to wire money, send cryptocurrency or read out gift card numbers are themselves signs of a scam.

Callback verification and code words

Callback verification beats every technical check because it does not depend on the quality of the fake. Hang up, dial the number saved in your phone, and if you cannot reach them, contact someone who can. Families and finance teams can also agree a code word in advance that a caller must give before any money conversation continues.

Listening for artefacts

Cloned speech can have flat emotion, odd pauses, mispronounced names or a metallic edge. Do not rely on this. A 2021 study by Müller and colleagues ran a game in which 472 people competed against a state-of-the-art detector over 14,912 rounds. Humans and the algorithm shared similar strengths and weaknesses, IT professionals had no advantage, and older participants were more susceptible. Artefacts are a reason for caution, not a method for clearing a call.

Detection services and their limits

Pindrop states on its site that it is independently validated at 99% accuracy, and its audio deepfake page says Pulse can flag a deepfake in two seconds; both are vendor figures. Resemble Detect claims up to 99.5% accuracy for its multimodal model, again self-reported. Hive offers audio detection scored in ten-second chunks.

The ElevenLabs AI Speech Classifier is different: it answers only whether a clip was made with ElevenLabs, uses the first minute of the upload, and the page itself says it does not reliably classify audio from the company's Eleven v3 model. A negative result means "not detectably ElevenLabs", nothing more. All of these return a probability, the kind of evidence discussed in the text detector accuracy thread.

For published recordings, provenance can help: the C2PA standard covers audio, so a signed Content Credential can establish a file's origin. See the C2PA thread for what that does and does not cover.

The legal backdrop

The US Federal Communications Commission has confirmed that the Telephone Consumer Protection Act's restrictions on artificial or prerecorded voices cover AI-generated voices, so such calls require prior express consent. That helps against robocall operations; it does not stop a one-to-one scam call.

Signs of a cloned voice, how much to trust each, and what to do

Sign	Reliability	Action
Urgent request for money by wire, crypto or gift card	High as a scam indicator, whatever the voice sounds like	Hang up, call back on a known number
Caller resists a callback or a code word	High	Treat as fraud until verified
Flat emotion, odd pauses, metallic tone	Low, humans and detectors both miss good fakes	Slow the call down, verify independently
Caller ID matches the person	Low, numbers are easily spoofed	Ignore it as evidence
Vendor detector score (Pindrop, Resemble, Hive)	Moderate, vendor accuracy figures are self-reported	Use to route calls for extra verification, not to decide alone
ElevenLabs classifier result	Only tells you about ElevenLabs audio	Do not read a negative as "real"

What that means in practice

• Verification is a process, not a sound. Callback on a known number and a shared code word defeat a perfect clone.
• Payment method is the tell. The FTC's list of hard-to-reverse channels is the most reliable sign you have.
• Detectors are for screening. Let a score add friction to a call, never let it clear one.