A finance worker in Hong Kong watched his CFO on a video call, then wired out $25.6 million. Every other person on that call was a deepfake. Learning how to spot AI scams stopped being optional the day that happened.
The $893 million problem
Here’s the scale, and it’s ugly. The FBI’s Internet Crime Complaint Center logged $893 million in reported losses from AI-powered scams in a single year, spread across 22,364 complaints. That was the first year the FBI gave AI its own section in the annual report. Meanwhile, FTC data shows Americans lost $665 million to phone fraud in just the first half of 2026.
And those are only the reported numbers. Most victims never file anything, out of embarrassment or plain disbelief. A writer named David Ralph Lewis put the vibe perfectly in a recent essay: scams used to be something filters caught. Now they’ve “escaped the spam filter and run wild.” You see it in the ads too. A Tai Chi program promising to get you ripped in 20 days. Glasses that adapt to “any prescription.” A USB stick offering free unlimited internet. All AI-generated, all paid for, all live on major platforms.
The scammers don’t need to fool you, though. They need to fool 1 in 10,000 people to make rent. So the question isn’t whether these fakes will reach you. It’s whether you’ll recognize one when it does.
The five scams that actually reach you
Not every AI threat matters to a regular person. These five do, because they all end with someone asking you for money or information. Knowing how to spot AI scams starts with recognizing the shape of each one.
Cloned voice calls
This is the big one. Researchers at Microsoft showed back in 2023 that a system called VALL-E could clone a voice from a 3-second audio clip. Anyone who has ever posted a video online has probably left enough audio lying around. Worse, one study found 70% of people can’t tell a 3-second clone from the real thing, and frequent AI users did no better than everyone else.
The pattern is almost always the same: a “family member” calls in a panic, often with crying or sirens in the background. In fact, about 92% of these calls use distress audio, because panic shuts down your critical thinking. The family emergency story accounts for roughly half of all voice-clone scams, with an average loss around $4,200 per victim. Seniors absorb 64% of the losses.
Deepfake video calls
Voice was step one. Video is step two, and the Hong Kong case is proof it works at scale. A fake CFO plus several fake colleagues convinced one employee to move $25.6 million. You’re not paranoid if you double-check a video call from “the office.” You’re paying attention. For suspicious clips, our guide on how to check if a video is AI-generated walks through a free way to test them.
AI-written phishing emails and texts
Generative AI killed the typo tell. Old scam emails were full of broken English, so your brain learned to spot them. Now the messages are flawless, and they’re personalized too. Lewis described getting a book-promotion email that echoed the exact phrasing from his own website. Someone had pointed ChatGPT at his page, dressed up as a flattering review. It was actually phishing for info or money.
So the modern question of how to spot AI scams in your inbox comes down to context, not typos. Texts are the main entry point now, according to the FTC, and fake invoices are the classic corporate version. We covered that angle in depth in our guide to spotting AI-generated fake invoice emails.
Too-good-to-be-true ads
AI image and video generators let anyone manufacture a product demo for something that cannot exist. The tells are physical, not technical. Sunglasses that fix your vision ignore how optics work. An air conditioner the size of a book can’t cool a room in 90 seconds, because thermodynamics doesn’t take requests. A company spending heavily to promote a $19 miracle gadget with a stock-look demo is not a company with a product.
Fake “officials”
Business impersonation alone cost Americans nearly $128 million in the first half of 2026. Government impersonation sits right behind it in the FBI’s data. The caller ID can be spoofed, the voice can be cloned, and the script is usually urgency plus authority. Your bank “flagging suspicious activity” is the most common flavor.
Here’s the cheat sheet version:
| Scam type | The tell | Your move |
|---|---|---|
| Cloned voice call | Urgency, distress audio, odd payment demands | Hang up, call back on the real number |
| Deepfake video call | Odd lighting, flat expressions, weird pauses | Ask something only the real person knows |
| AI phishing email or text | Perfect grammar, odd sender address, rushed link | Hover before clicking, verify separately |
| Fake product ads | Physics-defying claims, no real company info | Search the brand plus the word “scam” |
| Fake official | Pressure to act now, requests for secrecy | Hang up, contact the organization yourself |
The 7 checks that catch AI scams
You don’t need special software for most of this. You need seven habits, and each takes seconds once they’re automatic.
1. Hang up and call back. Use the number on your bank card or the family member’s saved contact, never the number that just called you. A real emergency survives a two-minute callback. A scam usually doesn’t.
2. Set a family safe word. Pick something random, nothing tied to your pet’s name or your birthday. If a “relative” can’t produce it, the call ends. Only about 18% of families have done this, which is honestly low given it takes five minutes.
3. Ask the unpredictable question. Clones respond to prompts, not memories. Ask what you ate together on some random Tuesday, or have the caller put a hand flat across their face. Early deepfakes struggle with the hand test, though they’re improving fast, so treat it as a signal, not a guarantee.
4. Watch the payment method. Roughly 62% of voice-scam payouts are demanded through wire transfers or Bitcoin ATMs, because those are hard to reverse. Gift cards are another classic. No legitimate organization asks for money that way. Ever.
5. Slow the story down. Urgency is the product. Crying in the background, “don’t tell mom,” “the deadline is in one hour,” all of it exists to keep you from thinking. So think anyway. Even five minutes of delay kills most scams, because the scammer is running dozens of targets and can’t afford to wait around.
6. Check the details that AI gets wrong. Look at the actual sender address, not the display name. Hover over links before clicking. For ads, reverse-image search the product photo and ask whether the thing physically makes sense. Our walkthrough of Chrome’s scam detection alerts adds a free automated layer on top.
7. Verify “official” contacts independently. The FCC ruled in 2024 that AI voices in robocalls are illegal without consent, so a “government” robot caller is already breaking a real rule. Hang up, look up the agency’s number yourself, and call that. Real agencies also don’t demand gift cards or crypto, which the FTC states plainly on its fraud pages.
One habit that beats everything
If you remember one thing from this whole guide, make it the callback habit. Voice cloning is cheap and video deepfakes are getting good, so you cannot trust your ears alone. What you can trust is a channel you control. The entire trick behind how to spot AI scams quickly is refusing to continue a conversation on the terms the caller set. Hang up. Breathe. Call back on the number you already have. That’s it.
It feels rude the first few times. Do it anyway, because the polite people are the ones getting drained.
Your 5-minute family setup
Tonight, do four things. First, pick a safe word with your parents and your kids, and explain the hang-up-and-call-back rule in one sentence. Second, remind everyone that voices on TikTok and Instagram are raw material for clones, since research attributes 58% of harvested voice samples to social media clips, so maybe skip posting your kid’s voice in public videos. Third, have one real conversation with the seniors in your family, because they carry nearly two-thirds of these losses and usually haven’t heard the 3-second-clone fact. Fourth, save two links on every phone in the house: the FBI’s complaint portal at ic3.gov and the FTC’s reporting site at reportfraud.ftc.gov.
If money already left your account, call the bank immediately anyway. Recovery odds are poor, under 2.5% by some estimates, but fast reports occasionally freeze transfers before they settle. Then file the reports. It takes ten minutes and feeds the data that eventually becomes enforcement.
That’s how to spot AI scams without living in paranoia. Seven checks, one safe word, one habit. The fakes are getting better every month, but the callback habit doesn’t care how good they get.