AI Voice Phishing Performs on Par With Human Scammers at a Fraction of the Cost

LessWrong reports on a large-scale study of AI-powered voice phishing attacks. Researchers tested four thousand one hundred participants against recordings of scams generated by six leading AI voice models—Sesame, ElevenLabs, Llama Full Duplex, Gemini, OpenAI AVM, and Play.AI. In the most emotionally manipulative scenario—a cloned voice impersonating a sister in distress—thirty-six percent of participants would or might comply with the scam. The overall compliance rate across all five scam types was sixteen point five percent. Participants couldn't reliably distinguish these AI voices from human callers, and familiarity with AI systems provided no advantage. Persuasiveness—the content and delivery of the message—proved to be the strongest predictor of compliance, far outweighing voice quality. That raises an unsettling economic reality: while human-operated voice phishing is unprofitable at US wages, AI-powered versions appear to be economically viable for several leading models.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton