A voice once offered something powerful in a telephone conversation: a sense that the person speaking was who they claimed to be. Artificial intelligence is weakening that shortcut. Modern voice clones can reproduce familiar vocal characteristics, leaving individuals and businesses confronting a problem that ordinary listening may not solve.
The warning became more immediate on October 19, when INTERPOL launched an awareness campaign about AI-enabled impersonation scams. The agency described fraud involving calls from supposed colleagues or relatives, alongside investment and romance scams.
Its message reaches beyond individual caution: financial systems and digital products need ways to establish identity without treating a convincing voice as proof.
What Listeners Actually Hear
A study published in Speech Communication by Eugenia San Segundo, Aurora López-Jareño, Xin Wang and Junichi Yamagishi tested 54 listeners—28 native Spanish speakers and 26 native Japanese speakers—using 80 voice samples, half artificially generated.
Participants achieved just 59.11% average accuracy when judging whether speech was natural or synthetic.
The researchers examined two languages, audiobooks and interviews, and whether listeners knew the speaker. Participants generally relied on intonation, rhythm, pauses, breathing and laughter rather than individual speech sounds.
Spanish listeners averaged 60.52% accuracy, compared with 57.60% for Japanese listeners. However, these findings concern specific experimental samples and should not be generalised to every language or voice-cloning system.
These cues matter because speech identity involves more than pitch or accent. Natural conversation carries patterns of timing, hesitation and expression. Generative models can imitate increasingly many of those characteristics, making familiar signs of authenticity less dependable.
A separate 2026 Scientific Reports experiment by Matthew Ivory, Carolyn McGettigan, Helen E. Nuttall and Sophie J. Nightingale asked 172 participants to classify real and AI-generated faces and voices.
Participants correctly classified 65.48% of faces but only 53.95% of voices. Crucially, confidence tracked face-classification accuracy but not voice accuracy.
That distinction is unsettling: someone can feel confident about a voice without being particularly good at determining whether it is synthetic.
Recognising Someone Is Not Verifying Them
Research published in Scientific Reports in 2025 helps explain the problem. Sarah Barrington, Emily A. Cooper and Hany Farid conducted listening experiments involving 604 participants and recordings derived from 220 speakers.
Participants judged an AI clone to be the same speaker as its real counterpart approximately 80% of the time. When explicitly asked whether a recording was real or generated, they identified synthetic voices correctly about 60% of the time.
These were different experimental tasks: recognising an identity and detecting synthesis. Their divergence illustrates why an apparently familiar voice cannot independently establish that a caller is genuine.
Earlier research published in PLOS ONE by Kimberly T. Mai, Sergi Bray, Toby Davies and Lewis D. Griffin reached a related conclusion.
Among 529 participants hearing English and Mandarin samples, listeners correctly identified deepfake speech 73% of the time. Providing examples beforehand improved recognition only slightly.
The accuracy rates differ across studies because stimulus quality, speakers, listening conditions and experimental tasks differ. They should not be combined into a universal likelihood of deception.
Why Voice Authentication Needs More Than Voice
For financial institutions and businesses, the issue is not simply whether synthetic speech sounds natural. It is whether an attacker can use a credible voice to persuade someone to disclose information, approve a transaction or bypass a security check.
INTERPOL warned that convincing AI-generated faces and voices can be created from relatively small quantities of online material.
Toshinobu Yasuhira, director of its Innovation Centre, cautioned that traditional warning signs are disappearing as generative tools improve.
Technical detection is necessary but difficult.
The peer-reviewed ASVspoof 5 research, published in Computer Speech & Language in January 2026, introduced a database with speech from approximately 2,000 people and attacks generated using 32 algorithms.
Its design incorporates varied recording conditions and adversarial attacks to evaluate speaker-verification and deepfake-detection systems more realistically.
That work highlights a product-design challenge: a detector that performs well on familiar recordings or attacks may struggle when microphones, compression, languages or synthesis techniques change.
Rebuilding Trust Through Verification
INTERPOL recommends independently contacting people through known channels and adopting multi-step verification for financial transactions or sensitive data transfers.
Those protections do not depend on hearing a flaw in someone’s voice.
For founders and security engineers, the practical implication is to separate a claim of identity from authorisation to act. A familiar voice might initiate a request, but high-risk actions should require independent confirmation, strong authentication and a traceable approval process.
That approach cannot eliminate every scam. It can, however, remove the most dangerous assumption from a digital interaction: that sounding human—or sounding like someone known—means the speaker is trustworthy.
As synthetic speech improves, trust must become something systems verify rather than something voices merely evoke.
