
Most AI scams are old scams: the relative in trouble, the boss who needs an urgent transfer, the online romance that turns into an investment tip, the shop with impossible prices. What AI has changed is the giveaways. The voice on the phone sounds like your son, the email has no typos, the face on the video call smiles and nods. So stop looking for flaws and check instead: call back on a number you already had, ask something only the real person knows, and look up any site yourself before you pay or sign in.
The FBI's 2025 Internet Crime Report, the first to give artificial intelligence its own section, counted 22,364 US complaints that mentioned AI, with losses of $893 million, and warns that "many victims do not realize the extent AI may be involved".
Three checks that work whatever the fake
In May 2025 the FBI warned that AI-generated content "has advanced to the point that it is often difficult to identify". So don't rely on catching a glitch:
- Call back on a number you already had. Hang up and ring the person or organisation on a number from your contacts, their official website or the back of your card, never one from the message or the call. Caller ID proves nothing: the FBI says scammers spoof legitimate phone numbers, names and credentials.
- Ask something only the real person knows. A family code word, or a detail that isn't on anyone's social media. A clone can copy a voice and a face, not a private memory. In July 2024 a Ferrari executive stopped a call from a convincing copy of the chief executive's voice by asking for the title of a book the real one had recommended days earlier, and the caller hung up, according to Bloomberg's report.
- Check the site behind any link yourself. Type the address of the bank or shop rather than tapping the link, or paste the link into a checker before opening it. A fake site can be built in minutes, but its domain still has to be registered, and that leaves a trail.
Then notice the shape of the request. Nearly every AI scam ends the same way: urgency, secrecy, and money or a code. "Don't tell anyone", "it has to be today" and "pay by gift card, crypto or wire" are the scam talking, whoever's voice it borrows.
Cloned voices: the call from someone you love
The FTC warns that a scammer needs only a short audio clip of your family member's voice, which can come from something posted online, and a voice-cloning program. Its advice: "Don't trust the voice."
The usual version is the distress call: a grandchild arrested abroad, a daughter in a car crash, a kidnapping. Hang up and call back on a number you know is theirs, or through another relative or their friends if you can't reach them. Treat the payment method as a verdict: wire transfers, crypto and gift card numbers are how these callers want paying, the FTC says, because the money is hard to get back.
Agree a family code word
The FBI recommends you create a secret word or phrase with your family to confirm who you're talking to. A few rules make it work:
- Pick something that can't be found online: not a pet's name, a school or a street you've posted about.
- Agree it in person or on a call, not in the family group chat, where a hacked account could read it.
- Ask the caller to say it rather than hinting at it. If they can't, hang up and call back on their usual number, whatever reason they give.
Deepfake video calls
Seeing a face is no longer proof either. Early in 2024, an employee in the Hong Kong office of Arup, the British engineering firm, received a message that appeared to come from the company's UK-based chief financial officer, about a secret transaction. Hong Kong police said the employee suspected phishing at first, then joined a video call with what looked and sounded like the CFO and several colleagues. All of them were deepfakes. The employee made 15 transfers totalling about US$25 million, and found out only after checking with head office. Arup confirmed it was the company in May 2024.
A live deepfake can still slip:
- The glitches the FBI lists: lag, a voice that doesn't quite match the face, unrealistic movement and, in pictures, distorted hands, teeth or eyes.
- Dodging a live test. The US Treasury's financial crime unit, FinCEN, has told banks that criminals using deepfakes to open accounts may claim repeated technical glitches or ask to switch to another channel when asked to prove who they are on camera. A caller whose video keeps failing just as you ask a question is doing the same thing.
But it was the video call that settled the Arup employee's doubts, so don't wait for a flaw. If a call ends with a request for money, a password or secrecy, close it and contact the person through a number or account you already had. Businesses should make that a rule: no payment and no change of bank details on the strength of a call or message alone, and a second person to approve large transfers.
AI-written phishing
Phishing emails used to give themselves away with clumsy English. The UK's National Cyber Security Centre judged in January 2024 that generative AI will make it difficult for everyone to tell whether an email or password reset request is genuine. In a study published in November 2024, emails that an AI agent wrote from public information about each recipient were clicked by 54% of the people who received them: as many as emails written by human experts, against 12% for generic phishing.
So judge the request, not the writing. A polished email that wants you to sign in, pay, change bank details, read out a code or download something deserves the same suspicion as a scruffy one. Check the sender's real address and where any link goes; our guide to phishing emails, texts and fake login pages covers both.
Fake shops, reviews and endorsements
Microsoft reported in April 2025 that with AI, fraudulent shops can be set up in minutes, complete with generated product descriptions, images, customer reviews, influencer endorsements and an invented business history. Celebrities get generated too: the FBI's annual report says investment scams use AI-generated videos and voices of celebrities, CEOs and other trusted figures in professional-looking endorsements on social media.
What still gives them away:
- The domain's age. A shop can invent its history but not its registration date, which ICANN's lookup shows for any domain. Of the sites verified.fyi has rated Dangerous where the registration date is known, 64% were less than a month old.
- Reviews somewhere else. Look the shop up on a review site you open yourself, and search its name with "scam" or "complaint".
- The face. A reverse image search catches a founder's photo stolen from someone else. A freshly generated face won't turn up anywhere, so no match proves nothing.
- The celebrity. Assume a famous face selling an investment or a miracle product in an ad is fake unless the celebrity's own accounts say otherwise.
Our guide to spotting a fake online shop goes through the rest.
Romance and investment scams with generated personas
AI makes the fake person cheaper to run at scale. The FBI describes criminals using AI to create fake social media profiles in large numbers, with realistic photos, forged IDs and "private" pictures, and AI-written copy for fake crypto investment sites.
In October 2024, Hong Kong police arrested 27 people over what they called deepfake love scams. The syndicate used AI-generated images to create attractive personas on social media, then steered victims into cryptocurrency investments backed by forged trading records, in scams totalling about HK$360 million. Many realised only when they couldn't withdraw their money.
The signs that matter more than the photos:
- Money for someone you've only met online. The FBI's advice is never to send money, gift cards or crypto to people you've met only online or over the phone, however long you've been talking.
- Investment talk from a new friend. A romance or a "wrong number" text that turns to a trading platform is the opening our guide to crypto investment scams describes.
- AI as the selling point. The US Commodity Futures Trading Commission warns that AI can't predict the future, and that promises of huge returns or 100% "win" rates from AI trading bots have been false.
- Fees to withdraw. If you must pay tax or fees before you can take money out, the profits were never real.
If you've already paid or shared something
- Call your bank on the number on your card. The sooner you do, the better the chance of stopping the money.
- Secure your accounts. If you gave a password, change it on the real site and anywhere you reused it. The FBI's rule is never to hand a two-factor code to anyone by email, text or messaging app.
- Keep the evidence: phone numbers, messages, payment details, crypto addresses, images and video.
- Report it. In the US, at ReportFraud.ftc.gov and the FBI's IC3. In England, Wales and Northern Ireland, to Report Fraud; in Scotland, to Police Scotland on 101. In the UK, forward scam emails to report@phishing.gov.uk and texts to 7726, and report scam ads to the ASA.
Check the site behind the link
Many AI scams end with a link: a payment page, an investment platform, a shop. That part you can check. Paste the address into verified.fyi and we collect 200+ facts about the site and score it from 0 to 100, including:
- how old the domain is, and whether a "family business since 1998" or thousands of reviews fit a domain registered weeks ago;
- whether the address imitates a well-known site, and whether testimonials show faces from services that hand out portraits for design mock-ups;
- whether a licence number the site quotes is on the FCA or ASIC register, or belongs to a firm that gave the regulator a different website;
- whether Google Web Risk, VirusTotal, Spamhaus or abuse.ch's threat feeds have flagged it, or it shares a payment account or crypto wallet with sites flagged for phishing or malware.
It can't tell you whether the voice on the phone is your daughter's; only calling her back can. But it can tell you a lot about where the call wanted to send you. The methodology explains each check, and our guide to telling if a website is legit covers what you can check by hand.