You’re at work. Your phone rings, and it’s your sister. She’s scared, she’s in trouble, and she needs money moved right now. It sounds like her. What do you do?
That’s the question behind a lot of the AI voice scam worry, and a recent study gives us some actual numbers on it. They’re not as scary as the headlines, but they point to something useful you can do this week.
What the study found
Researchers affiliated with Harvard University, Meta and other organisations ran an experiment with 4,100 U.S. adults. It was reported by Japanese outlet ITmedia and summarized by DigitalToday. They compared six AI voice models against real human voices, across five scam scenarios:
- pretending to be a card issuer
- pretending to be Gmail customer service
- asking for donations
- posing as police demanding bail for a relative
- pretending to be a sister in danger who needs an urgent transfer
Participants listened to call recordings or read transcripts, then said whether they’d go along with the request.
Across all scenarios, an average of 16.5 percent said they would comply or found it hard to decide. That’s about 1 in 6. In the scenario that used ElevenLabs’ voice-cloning technology to pose as a sister in danger, the number rose to 36.1 percent.
The study found that emotional, personal stories were more likely to get a yes than the usual “customer service” approach.
The voice isn’t the main weapon. The story is.
This is the part I’d want you to remember. The researchers concluded that a story that creates urgency and sympathy matters more than how natural the AI voice sounds.
That matches how scams work in general. A scammer wants you to act before you think. A panicked relative, a police officer, a bail amount and a deadline all push you to skip the normal checks. A convincing voice helps, but the pressure does most of the work.
People also aren’t great at spotting the fakes. In the study, 70.3 percent of participants correctly identified an AI voice. But 67.4 percent wrongly judged a real human voice to be AI. Frequent users of generative AI or voice assistants weren’t clearly better at telling the difference.
So “I’ll know it when I hear it” isn’t a plan. Even if you can sometimes tell, you can also be wrong the other way and hang up on a real person who needs you.
Why this matters for the cost of scams
The study also did a rough economic estimate. Using U.S. labour costs as a benchmark, a human making scam calls would lose about $27 per hour. Some AI models, by contrast, could produce an expected profit of $1 to $3 per hour. The researchers warned that as the technology spreads, scam calls could be automated and the cost of mass dialling could drop.
In plain terms: calling thousands of people gets cheaper, so more people get called.
What this study does not say
Be careful with the 16.5 percent. The study measured survey answers, not actual fraud losses. Participants reviewed recordings or transcripts. They weren’t getting a surprise call at dinner. The source is clear that 16.5 percent should not be read as a real-world success rate. It also covered U.S. adults, not Canadians specifically.
So no, this isn’t proof that one in six of us will send money. It does show that the family-emergency script is the one to prepare for.
How to tell if this could affect you
Honestly, if you have a phone and family, it could. A few things make the risk higher in practice:
- You’d feel obligated to help a relative immediately, without question.
- Your family members are used to sending you or receiving money by quick transfer.
- Nobody in your family has agreed on how to verify a stressful call.
That last one is fixable in about ten minutes.
What to do today: a code word and a callback routine
These are standard, widely recommended steps. They don’t rely on your ear, which is the point. I’d suggest running through them with your family this week as a short practice drill, so they feel normal before you need them.
1. Pick a family code word
- Choose a word or short phrase that isn’t tied to anything on social media: not a pet’s name, a street, a school or a birthday.
- Make it something odd and easy to remember. Two unrelated words work well.
- Share it in person or on a call you started yourself. Don’t send it in a text thread you wouldn’t want a stranger to read.
- Agree on the rule: anyone asking for money or urgent help on a call has to give the word. No word, no money, no exceptions, even if they sound right.
- Tell older relatives and teenagers too. Everyone needs to know the rule.
2. Use a callback routine
A code word can be forgotten in a panic, so add a second layer that doesn’t depend on memory of the word.
- Stop. Say you’ll call right back. A real family member in trouble will accept that. A scammer will push back or say there’s no time. That pushback is a warning sign.
- Hang up.
- Call the person on the number you already have for them, from your contacts, not the number that just called you and not one the caller gave you.
- If they don’t answer, try another family member or a friend who’d know where they are.
- Don’t move any money until you’ve reached the real person.
The same goes for the “police officer” call asking for bail. Police don’t need you to hang up and send money in the next ten minutes. If you’re worried it’s real, hang up and contact the police service yourself using a number you look up independently.
3. Write it down
Put the code word and the callback steps on a card or in a shared note your family can find. Under stress, people forget things they “already know.”
Signs something already went wrong
- You sent money and then couldn’t reach the relative, or you reached them and they knew nothing about it.
- The caller asked you to keep it secret or told you not to call anyone else.
- You shared a code word, banking details or a verification code during the call.
If any of that happened, call your bank or payment provider right away and ask what can be stopped or reversed. The faster you move, the better your odds. Then report it. In Canada, the Canadian Anti-Fraud Centre and your local police are the places to do that. If you gave out passwords or codes, change those passwords from a different device and turn on two-step sign-in where you can.
One more thing
If someone in your family falls for it, don’t make it a lecture. The study suggests emotional scripts work on a lot of people, and the point of the code word is that nobody has to rely on being clever under pressure.


