**Most people can't reliably tell an AI voice from a human one on a phone call — and that's not actually the risk it sounds like, because voice realism barely predicts whether someone falls for a scam.** Harvard Kennedy School and Meta researchers found the script's persuasiveness drives compliance, not how human the caller sounds.
Can People Actually Tell an AI Voice From a Human One?
Barely. Harvard Kennedy School, Harvard's engineering school, and Meta ran a controlled study of 4,100 U.S. adults — nationally representative, built on six commercial AI voice systems (Meta's own Llama Full-Duplex, OpenAI's Advanced Voice Mode, Google Gemini, Sesame, Play.AI, and ElevenLabs) plus real human callers as the control (Harvard Kennedy School/Meta, 2026). Participants heard recorded phishing-style calls and rated them on sentiment, persuasiveness, trustworthiness, and human-likeness; researchers then checked which trait actually predicted whether someone agreed to a risky request. Human callers pulled a 21.4% compliance rate. AI voices pulled 15.4% — a real gap, but not one reliable enough to use as a safety test on any single call (Harvard Kennedy School/Meta, 2026).
What Actually Predicts Whether Someone Falls for the Call?
Not the voice. The same study found persuasiveness was the strongest predictor of compliance — each step up the persuasiveness scale raised the odds of going along with the request by a factor of 2.58. Human-likeness, once persuasiveness and the other traits were accounted for, stopped predicting compliance at all (p = .222) (Harvard Kennedy School/Meta, 2026). A well-paced, well-written script beat a convincing voice every time. The riskiest category tested — a caller posing as a distressed relative — pulled a 36.1% response rate, more than five times the rate for a generic account-support scam (Harvard Kennedy School/Meta, 2026).
| Common assumption | What the research actually found |
|---|---|
| "I can tell when a voice is AI-generated." | Compliance only dropped from 21.4% (human) to 15.4% (AI) — a real gap, not a reliable test on any one call |
| "A robotic-sounding voice means it's probably a scam, or probably safe." | Human-likeness had no independent effect on compliance once persuasiveness was controlled for (p = .222) |
| "Urgency would make me more suspicious, not less." | The distressed-relative scenario hit 36.1% compliance — the highest of any category tested |
| "Legitimate businesses don't sound scripted." | Human and AI callers read identical scripts in the study — the script moved the outcome, not who or what read it |
Why This Is a Live Problem for Businesses Right Now
Because adoption and fraud are climbing together. Sinch's 2026 survey of decision-makers at North American enterprises found 64% already have AI voice agents in production and 97% plan to increase that investment this year (Sinch, 2026). At the same time, the FTC's imposter-scam data for 2025 shows $3.5 billion in reported losses — nearly $1 billion of it from callers impersonating a bank, business, or government agency, with reported losses up nearly three-fold since 2020 (FTC, 2026). Every business switching on an AI phone line this year is stepping into that same environment, where a customer's go-to test for a call — does this sound human — has already been shown not to work.
When a Human Should Make the Call, Not AI
Automate the wrong kind of call and a business doesn't just risk a bad experience — it risks training its own customers to comply with the exact pattern a scammer runs. Any outbound call asking someone to confirm payment details, move money, or "verify" account information is the highest-risk category to hand to AI, precisely because the research shows persuasive phrasing works regardless of who — or what — delivers it. A booking confirmation, an appointment reminder, or a "press 1 for a person" routing call carries none of that risk. A call asking a customer to read back a card number does, whether a human or an AI reads the script.
How to Keep an AI Receptionist From Getting Mistaken for a Scam
- Never ask for payment details, account verification, or money movement on an outbound call — route anything with a dollar sign to a callback on a number the customer already has.
- Publish one real callback number and tell callers to use it to verify anything urgent, instead of trusting whoever just called them.
- Keep a human one request away on every call — it's the single feature that most closes the gap between a customer using AI and a customer trusting it.
- Disclose the AI upfront where state law requires it. It won't stop a scammer from skipping that step, but it keeps a business the one that followed the rule when a customer checks.