How criminals are using AI: inside the new scam playbook

AI hasn't invented new crimes. It has upgraded old ones: the fake bank call, the "Hi Mum" message, the too-good-to-be-true investment. Here is what criminals are actually doing with AI, drawn from police data, court cases, bank reports and peer-reviewed research, and the habits that still beat every one of them.

• 51% of spam emails were written by AI by April 2025, up from 16% a year earlier¹
• £221.5m lost to investment scams in the UK in 2025, a record, driven by fake adverts and social media posts²
• 395% rise in UK fraud reports where victims said AI was involved (956 reports in 2025/26)³
• 73% is how often listeners correctly spotted a fake voice in a lab test. In other words, about one fake in four got through⁴

How a scam works, and where AI fits in

Almost every scam follows the same four steps. The criminal finds a target, hooks them with a message, call or advert, keeps them talking until they trust the story, and then gets them to pay. AI now helps at every step:

• Finding: sorting through leaked data and public social media to pick targets and personalise the approach.
• Hooking: writing fluent messages in any language, and making fake adverts with real celebrities' faces and voices.
• Keeping you talking: chatbots and translation tools that let one criminal run dozens of conversations at once, and AI photos of people who don't exist.
• Getting you to pay: a cloned voice of a relative, or a deepfake video call with "your boss", at the moment that matters.

The scale of the underlying problem is huge. UK banks confirmed 4.06 million fraud cases in 2025 and £1.28bn stolen. Of that, £576.4m was lost in 248,070 scams where victims were tricked into sending money themselves.² Two in three of those scams started online.² How much of it now involves AI is a hard question, and I've tried to answer it in a separate piece with a full model. This article is about the how.

1. Perfect messages, sent by the million

For years the standard advice was to look for spelling mistakes and clumsy English. That clue is disappearing. Researchers from Columbia University, the University of Chicago and the email security firm Barracuda ran nearly half a million malicious emails through an AI detector. They found that at least 51% of spam was written by AI by April 2025, up from 16% a year earlier. For targeted "business email compromise" attacks, where criminals pose as a boss or supplier, the figure was at least 14%, up from 8%.¹ The authors call these conservative floors; the true share is probably higher.

Does AI-written phishing work? In a study at Harvard, fully automated AI phishing emails, researched and written by AI, got 54% of recipients to click a link. That was the same as emails written by human experts, and far above the 12% for ordinary phishing.⁵ It was a small study of about 100 volunteers, but the direction is clear. IBM's security team found that AI produced a convincing phishing email in five minutes; their human experts usually take about 16 hours.⁶

2. Celebrity deepfake adverts

Investment scams cost UK victims £221.5m in 2025, the highest figure ever recorded and 38% of all money lost to scams. UK Finance says they are "being driven by fraudulent adverts/social media posts".² Many of those adverts now use deepfakes: AI-generated video of a well-known face, usually a trusted money expert, a TV presenter or a politician, appearing to recommend an "AI trading platform".

Martin Lewis has been a favourite target since 2023, when he called a fake video of himself "frightening".⁷ In 2025 a Guardian investigation into a call centre in Georgia found that deepfake and fake-news adverts featuring Lewis, Zoe Ball and Ben Fogle had helped take around $35m from about 6,000 people, roughly a third of them in the UK.⁸ The individual cases are heartbreaking. A Surrey man lost £20,000 of his life savings after watching a deepfake of Lewis.⁹ In September 2026 police in Northern Ireland reported a victim who lost £250,000 after an AI video of "a well-known personality from the financial sector".¹⁰

The UK's new national fraud reporting service, Report Fraud, says investment fraud is the biggest single category among reports where victims identified AI (37%), and the most common age group of those victims was 60 to 79.³ Ofcom found that among people who believed they'd seen a deepfake, 45% said it was a scam advert, rising to 54% of over-55s.¹¹

3. Cloned voices of people you know

The "Hi Mum" scam used to arrive as a text from an unknown number. Santander's fraud team now reports criminals using AI to make voice notes that sound like the child, to make the message more believable.¹² Voice cloning is cheap and quick. Researchers at Queen Mary University of London made clones with a commercial tool from about four minutes of recordings; listeners judged the clones to be human 58% to 70% of the time, about as often as real voices.¹³ In a 2023 UCL lab test, listeners spotted fake speech 73% of the time, and training barely helped.⁴

In a 2024 survey for Starling Bank, 28% of UK adults said they had been targeted by an AI voice-cloning scam in the past year, and 46% hadn't known such scams existed.¹⁴ That 28% is people's own belief, and some will have mistaken ordinary scam calls for AI. But the risk is real. National Trading Standards has warned of criminals using fake "lifestyle survey" calls to record people's voices, then using clones of those voices to set up direct debits.¹⁵

4. The fake boss on a video call

The most famous case is Arup, the engineering firm. In January 2024 a finance worker at Arup's Hong Kong office made 15 transfers totalling HK$200 million (about US$25m, roughly £20m at the time) after a video call where the 'CFO' and colleagues were deepfakes; police described it in February 2024 and Arup confirmed it in May 2024, saying that "fake voices and images were used".¹⁷

Others have been targeted the same way. Criminals set up a fake WhatsApp account and a Teams meeting using a voice clone of Mark Read, then chief executive of advertising giant WPP; the attempt failed.¹⁸ A Ferrari executive stopped a cloned-voice call from "the CEO" by asking which book he had recently recommended.¹⁹ In Italy, reports in September 2026 said criminals used an AI clone of a lawyer's voice to help steal about €95m from Fideuram, part of the banking group Intesa Sanpaolo; about €53m was recovered.²⁰

5. Romance with a face that doesn't exist

Romance scams cost UK victims £39.2m in 2025, up 23%.² AI helps at every stage: a perfect profile photo of someone who doesn't exist, fluent messages in any language, and now video. In France, a woman lost €830,000 over about 18 months to a scammer who sent AI-generated photos and videos of "Brad Pitt".²¹ In Hong Kong in October 2024, police arrested 27 people in a romance syndicate that used face-swapping and voice-changing deepfakes to lure victims into a fake crypto platform; losses were HK$360m, about £35m.²²

Much of this is now run on an industrial scale from scam compounds in South-East Asia, often staffed by people who have been trafficked. The UN's crime agency estimates scam losses across East and South-East Asia, Australia and New Zealand at $88bn to $114bn in 2025, and says generative AI has "dramatically reduced the technical barriers and workforce requirements" for this kind of fraud.²³ The AI companies' own reports show what that means in practice. OpenAI says the scam networks it banned in 2025, from Cambodia, Myanmar and Nigeria, mainly used AI "as a scaling and efficiency tool": translating, writing scripts and even running staff rotas.²⁴ In September 2026 the AI company Anthropic reported shutting down a network of more than 20 dating apps in which over 4,700 AI personas chatted with at least 25,000 people in just two weeks. Three in four of the "people" on those apps were AI.²⁵

In the UK, the share of romance fraud reports that mention AI is still small but rising: from 0.5% to 0.8% a month in late 2025 to 1.5% to 2.0% in 2026.³ That only counts victims who realised AI was involved.

6. Fake shops, fake products, fake reviews

Purchase scams are the most common scam of all: 175,809 cases in the UK in 2025, 71% of all cases where people were tricked into paying.² Lloyds Bank says more than 70% of the purchase scams it sees start on Meta's platforms.²⁶ AI makes it easy to build a convincing shop in an afternoon, with polished product photos, descriptions and glowing reviews. When Santander showed people adverts for products that didn't exist, 74% failed to spot a fake padel-bat advert and 71% failed to spot an AI-generated dog feeder.²⁷ Fake reviews have been banned in the UK since April 2025, including reviews generated by software, and the Competition and Markets Authority now has powers to fine businesses directly.²⁸

7. Fake AI tools and "trading bots"

AI itself has become the bait. "Quantum AI" and similar names have been attached to fake trading platforms that promise an AI will grow your money automatically; Which? has repeatedly found deepfake adverts promoting them on YouTube and social media.²⁹ Other criminals advertise fake "AI video generators". Google's Mandiant team found one group running more than 30 such websites, promoted by thousands of Facebook and LinkedIn adverts that reached over 2.3 million people in the EU. Whatever you typed in, the site downloaded malware that stole passwords and card details.³⁰

8. Faked faces to open bank accounts

Scam money has to land somewhere, usually in a "mule" account opened with a stolen or invented identity. Banks check new customers with a selfie or short video, and criminals are using deepfakes to get past those checks. The identity firm Entrust says deepfakes now account for one in five biometric fraud attempts it sees.³¹ Sumsub, another identity firm, says deepfake attempts in the UK nearly doubled in 2025 (+94%).³² Both sell detection services, so treat their numbers as indications rather than official statistics. But the trend is clear enough that UK Finance's 2026 fraud report includes a warning, in a section written by two of its fraud-prevention partners, about "selfie injections" letting fraudsters bypass ID checks.²

9. Fake images used for blackmail

The cruellest use is sextortion: criminals threaten to share intimate images unless the victim pays. The National Crime Agency has warned schools that offenders may use AI-generated or digitally altered images of the victim, so the victim doesn't need to have shared anything at all.³³ Victims are mostly teenage boys. Since February 2026 it has been a criminal offence to create, or ask someone to create, a sexually explicit deepfake of an adult without their consent, and since June 2026 it has also been an offence to make or supply "nudification" tools.³⁴

What AI hasn't changed, and how to use that

Every tactic above changes what you see and hear. None of them changes what the criminal needs you to do. They still need you to:

1. Act quickly, before you can think or talk to anyone.
2. Keep it secret: "don't tell the bank", "don't tell Mum".
3. Use their route: their link, their number, their app, their "safe account".
4. Move money, or hand over a code you'd normally never share.

So the defence is not to get better at spotting fakes. Researchers keep showing that people can't reliably tell the difference, and the fakes will only improve. The defence is to check by a different route before you act:

• Hang up and call back on a number you already have: the one on your card, your contacts list or the company's own website.
• Dial 159 to reach your own bank's fraud team safely. It connects customers of more than 99% of the UK's retail bank current accounts with their bank.³⁵
• Use a family safe word for any urgent request for money.
• Your bank and the police will never ask you to move money to a "safe account", or to hand over a one-time passcode.
• Report it: texts to 7726, emails to report@phishing.gov.uk, and fraud to Report Fraud (reportfraud.police.uk or 0300 123 2040), which replaced Action Fraud in December 2025.³

This idea, that we need to stop teaching people to spot clues and start teaching them to check by a second route, is the subject of my own research paper, which looked at the scam advice of 45 UK organisations.

Every figure was checked against its original source. If you spot an error, please email enquiries@jameshayesconsulting.com and I'll correct it.

If you run a club or group in Canterbury, Whitstable, Herne Bay, Faversham or Thanet, I give a free talk on AI and the new scams, with time for questions and no sales pitch. Ring 07500 511462.

Sources

1. Hao, Tran, Rideout et al., "Do Spammers Dream of Electric Sheep?", ACM Internet Measurement Conference 2025. 239,145 spam and 242,413 business email compromise emails sent to Barracuda customers, February 2022 to April 2025. Several authors work for Barracuda.
2. UK Finance, Annual Fraud Report 2026 (covering 2025), June 2026. Totals, losses and cases by scam type, and where scams start (p.34). The AI and "selfie injection" wording is in a foreword by BioCatch and Nasdaq Verafin.
3. City of London Police, Report Fraud Annual Assessment FY2025-26, September 2026, and the press release. Counts are reports where the victim identified AI, so they understate AI's role.
4. Mai, Bray, Davies and Griffin (UCL), "Warning: Humans cannot reliably detect speech deepfakes", PLOS One, August 2023. 529 participants.
5. Heiding, Lermen, Kao, Schneier and Vishwanath, "Evaluating Large Language Models' Capability to Launch Fully Automated Spear Phishing Campaigns", 2024. 101 participants in four groups.
6. IBM X-Force, "AI vs. human deceit", October 2023.
7. TechCrunch, "Martin Lewis deepfake scam ad", 7 July 2023.
8. The Guardian investigation into a Georgia-based call centre, March 2025, as summarised by commonspace.eu.
9. Surrey Police and Crime Commissioner, "Surrey victim loses life savings to deepfake video", February 2026.
10. PSNI statement reported by the Irish News, 7 September 2026.
11. Ofcom, "Deepfakes: demean, defraud, disinform", July 2024. Poll of UK internet users aged 16 and over, June 2024.
12. Santander UK, quoted in "WhatsApp scam known as 'hi mum' evolving at 'breakneck speed'", The Independent, 5 May 2025.
13. Lavan, Irvine, Rosi and McGettigan (Queen Mary University of London), "Voice clones sound realistic but not (yet) hyperrealistic", PLOS One, September 2025.
14. Starling Bank, Safe Phrases campaign, September 2024. Mortar Research, 3,010 UK adults, August 2024.
15. Royal Borough of Greenwich Trading Standards, passing on a National Trading Standards warning, "Phone scam uses AI to clone victims' voices", February 2026.
16. FBI, public service announcement on generative AI fraud, 3 December 2024; Police Scotland, "Voice cloning scams".
17. CNN, "Arup revealed as victim of $25 million deepfake scam", 16 May 2024.
18. The Guardian, "CEO of world's biggest ad firm targeted by deepfake scam", 10 May 2024.
19. Bloomberg, "Ferrari narrowly dodges deepfake scam", 26 July 2024, reposted copy.
20. Reuters, "AI messaging scam costs Italy's top bank Intesa millions, sources say", 25 September 2026. Based on two unnamed sources; the banks declined to comment.
21. France 24, "How an AI Brad Pitt conned a woman", 16 January 2025.
22. The Standard (Hong Kong), "Police arrest 27 for deepfake love scams totaling $360m", 14 October 2024.
23. UNODC, Transnational Organized Crime Threat Assessment for South-East Asia 2026, July 2026. The loss range covers all scams in the region, not only AI-enabled ones.
24. OpenAI, "Disrupting malicious uses of AI", October 2025.
25. Anthropic, "Detecting and countering misuse of AI: September 2026", 10 September 2026.
26. Lloyds Banking Group, "Top five purchase scams of 2025", December 2025.
27. Santander UK research on AI shopping scams, reported by FinTech Global, 3 December 2025. Sample size not published.
28. Competition and Markets Authority, fake reviews guidance (CMA208), April 2025.
29. Which?, "Which? finds deepfake investment scams on YouTube", October 2025, and "Quantum AI scams warning", March 2026.
30. Google Cloud / Mandiant, "Text-to-Malware: How Cybercriminals Weaponize Fake AI-Themed Websites", May 2025.
31. Entrust, 2026 Identity Fraud Report, press release, November 2025. Global data, September 2024 to September 2025. Entrust sells identity verification.
32. Sumsub, Identity Fraud Report 2025-2026, press release, November 2025. Sumsub sells identity verification.
33. National Crime Agency, "NCA issues urgent warning about sextortion", April 2024.
34. Data (Use and Access) Act 2025, section 138, in force 6 February 2026; Crime and Policing Act 2026, section 99, in force 29 June 2026 (Home Office Circular 004/2026).
35. Stop Scams UK, 159.