Can you detect an AI voice?
Verify the person.

Listening can tell you what you heard. It cannot reliably prove who spoke.

Modern speech synthesis can reproduce a person’s voice from recordings. A call may also be compressed, noisy, or interrupted, which makes small audio details hard to judge. A pause, an odd word, or a smooth delivery may raise a question; none is a dependable test of identity.

Two research studies illustrate the limit. Barrington, Cooper, and Farid tested cloned voices based on 220 speakers with 604 listeners. Participants did not consistently identify the clones. The researchers used one voice-cloning service and a U.S. English-speaking sample, so the result is not a score for every model, language, or real-world call. A separate 1,276-person study across synthetic images, audio, video, and audiovisual clips found that people struggled to distinguish synthetic from authentic media by perception alone.

The practical conclusion is narrow: do not make a high-stakes decision from how a voice feels. Verify the identity and the request independently.

Break the pressure loop before you investigate.

Scam calls often create urgency: someone is hurt, a payment cannot wait, or an account will be closed. The request may sound like a family member, a manager, or a supplier. Pressure is a reason to pause, not a reason to skip verification.

1. End the current call. Do not use a callback number, link, or contact detail supplied by the caller.
2. Start a new conversation. Call a saved number you already know belongs to the person, or use a separate channel you have used before.
3. Check the exact request. Confirm the amount, account, deadline, and reason independently. For a workplace request, use the directory or approval process your team already relies on.
4. Do not send secrets to prove yourself. Do not read out a password, one-time code, recovery phrase, or payment credential on the caller’s instruction.
5. If you cannot verify it, wait. Contact another trusted person or the organization through its official contact route.

The U.S. Federal Trade Commission gives the same core advice for family-emergency voice-cloning calls: call the person at a number you know is theirs, and try another family member or friend if you cannot reach them. That guidance is practical anywhere; its reporting channels are specific to the United States.

A real voice can still carry a false story.

Authenticity and truth are separate questions. A real recording may be edited, old, or shared with a misleading caption. A cloned voice may repeat a real person’s words but place them in a new situation. Verification should cover both who made the statement and whether the claim is supported.

This is the same discipline used for images: trace the source, preserve context, and separate provenance from the claim attached to a file. Read the image verification guide ↗.

Metadata can add context. Its absence does not settle the question.

Some supported audio carries a detectable watermark. OpenAI’s content-provenance documentation describes checking for SynthID in supported audio files. A detected watermark is a useful signal about the file; a “not detected” result does not rule out AI generation or editing. Other generators may use different signals, and file handling can affect what survives.

Signed provenance systems such as C2PA can make certain creation and edit records tamper-evident when they are present and supported. The standard describes those records as trust signals, not a verdict that the content or its accompanying story is true. Check what a verifier actually inspected, who signed the record, and whether the details match the file and context you have.

Do not upload a private conversation to an unfamiliar detector simply to get a label. First consider whether the recording includes someone else’s voice, confidential work, or personal information, and review the verifier’s data handling.

A classifier answers a narrower question than “can I trust this?”

An audio detector estimates whether a file resembles the synthetic speech its method was built and tested to recognize. Results can depend on the generator, language, clip length, recording path, compression, and the evaluation data. A score does not establish the caller’s identity, prove the story, or authorize a payment.

If you use a detector in a review workflow, record the tool and version, keep the original file where lawful and appropriate, and compare its result with independently obtained evidence. Do not treat an automated result as a verdict about a person.

For a related question about text, see what AI detector scores can and cannot tell you ↗. The same caution applies: a score needs a task, a known test population, and context.

Move to the trusted channel and limit further exposure.

If you shared a code, payment detail, or account information, contact the relevant bank or service using its official app or a number from its website or card. Change affected credentials from a trusted device, revoke exposed sessions or recovery options where available, and tell the real person or organization through a known route. Keep the original message and any transaction records for reporting.

If the situation involves money, workplace access, or someone’s safety, tell the appropriate people promptly. Avoid forwarding the clip as proof until its source and context are understood; sharing can expose private voices or amplify a false claim.

Understand the system behind the voice.

For the technical path from microphone to generated reply, read how voice AI works ↗. For video, follow the guide to source and context checks ↗. For a broader view of model behavior, continue to why fluency is not evidence ↗ or explore the AI field guide ↗.