How many scams now use AI? A model for the UK

Nobody in the UK officially counts how many scams involve AI. Depending on how you measure it, the answer is less than 1% or more than half. I built a simple, open model to find a sensible middle, using the UK's own fraud figures and the best evidence on how criminals use AI. My best estimate is that around one in five UK payment scams in 2025 involved AI somewhere, and they accounted for about a third of the money lost.

• 0.17% of UK fraud reports in 2025/26 were ones where the victim said AI was involved¹
• About 1 in 5 UK payment scams in 2025 probably involved AI in some way: my central estimate is 22%, roughly 55,000 cases
• About £190m was lost to those scams, around a third of all scam losses (my estimate)
• About 1 in 14 scams probably used a fake face, voice or image, but those cases account for about 13% of losses, roughly £74m (my estimate)

1. Four ways to count, four very different answers

Ask "how many scams use AI?" and you'll find figures that seem to contradict each other. They don't. They measure different things.

• What victims report to police. The UK's new national service, Report Fraud, which replaced Action Fraud in December 2025, counted 956 reports in 2025/26 where the victim identified AI as a factor. That was up 395% on the year before, but still only about 0.17% of all reports.¹ The FBI's first count of complaints that mention AI found 2.2% of complaints and 4.3% of losses in 2025, and 7.3% of losses to investment fraud.²
• What victims say when asked. In a 2026 Gallup survey of 5,173 American adults, people who had lost money to a scam in 2025 said 12% of those scams involved AI or deepfakes.³
• What researchers measure directly. Running nearly half a million malicious emails through an AI detector, researchers found that at least 51% of spam and at least 14% of targeted "business email compromise" emails were written by AI by April 2025.⁴

The first three depend on the victim noticing AI. The last measures one channel, email, very precisely. The truth for all scams is somewhere in between.

2. Why victims can't tell

Most of what criminals do with AI is invisible to the person on the receiving end. An AI-written text message looks exactly like one typed by a person. A chat with a "romance" partner who is really a chatbot, or a scammer reading from AI-written scripts, feels like a chat with a person. Even a cloned voice is judged to be human more often than not.⁵

The organisations collecting the figures say so themselves. Report Fraud says the true extent of AI involvement is unknown "as victims may be unaware of its use".¹ The FBI notes that investment losses overall were many times larger than those with a reported AI link, "demonstrating that many victims do not realize the extent AI may be involved".² And the Alan Turing Institute's 2025 review of AI and online crime pointed out that UK data separating AI-enabled fraud losses from the rest doesn't exist.⁶

So the victim-reported figures are a floor, not an estimate. To get closer to the real picture, you have to start from the scams themselves and ask where AI is likely to be used.

3. How the model works

I started with the most reliable UK count of scams: cases confirmed by banks and payment firms, published by UK Finance. In 2025 there were 248,070 "authorised push payment" scams, where victims are persuaded to send money themselves, with £576.4m lost.⁷ UK Finance splits these into eight types. For each type I set a range for the share of cases where AI was used anywhere in the scam the victim saw: writing or translating messages, running chats, making fake images, cloning voices or faking video. I gave each range a low, a most likely and a high value, based on the evidence for that type of scam.

Why these ranges

• Investment and romance scams get the highest ranges. Investment scams are "driven by fraudulent adverts/social media posts", which increasingly use deepfakes of trusted faces.⁷ Both types are often run from scam compounds that use AI to translate and script conversations, and to run fake profiles at scale.⁸⁹ They are also where victims most often notice AI: 3% of UK investment-fraud reports and up to 2% of romance-fraud reports flagged it, the highest shares of any type.¹
• Invoice and CEO fraud are mostly email-led, so I anchored them to the measured 14% of targeted business emails written by AI, rising over time,⁴ with some room for voice and video clones of bosses.
• Purchase scams get a low range. Many are simple: a fake listing on a marketplace or a social media advert. AI makes the text and photos easier, but plenty of these scams need no AI at all.
• Bank and police impersonation get the lowest range. These are mostly live phone calls from human callers. AI plays a part in opening text messages and some automated calls; Europol reports voice bots now screening victims "at industrial volume".¹⁰

The model then draws 100,000 random combinations from these ranges and adds them up, weighting each type by its number of cases and its losses. It assumes that if criminals have adopted AI faster than I think in one type of scam, they probably have in others too. Running it with each type treated independently gives almost identical results.

4. What the model says

• Any AI involvement: about 22% of cases, roughly 55,000 scams. In 8 out of 10 simulations the figure was between 16% and 30%.
• Money lost: about 33% of losses, roughly £190m, with a range of 26% to 41%. AI features more in the bigger scams, investment above all, so its share of losses is higher than its share of cases.
• Fake faces, voices or images: about 7% of cases (roughly 16,000) but about 13% of losses (roughly £74m), with ranges of 4% to 9% and 9% to 17%. Deepfakes are rarer than AI-written text, but they cluster in the most expensive scams.

These estimates sit well above what victims report, and below the share of spam written by AI. That is what you'd expect: AI is mostly invisible to victims, and many scams don't start with an email at all.

5. What drives the answer

I changed one assumption at a time to see which matters most. The answer is surprising: it isn't deepfakes.

• For the number of cases, it's purchase scams. They make up 71% of all cases, so the overall figure swings from 13% to 34% depending only on whether purchase scams are at the low or high end of their range. Every other type moves the total by 4 percentage points or less. The biggest unknown in UK scams isn't the Hollywood-style deepfake. It's how many of the fake sellers on social media and online marketplaces are now using AI to write their listings and make their photos.
• For money lost, it's investment scams. They account for 38% of losses, and moving them from the low to the high end of their range shifts AI's share of losses from 23% to 41%.
• If all my ranges are too high by half, AI was involved in about 10% of cases. If they are all 50% too low, it's about 30%. Either way it's far more than victims report.

6. What the model can't tell you

This is a model, not a measurement, and it has clear limits:

• The ranges are judgements. They are anchored to published evidence, but no dataset measures AI's role in each type of UK scam. Different reasonable judgements would give different answers, which is why I've published every one of them above.
• "Involved" doesn't mean "caused". Most of these scams existed before AI. OpenAI found that the scam networks it banned in 2025 used AI mainly "as a scaling and efficiency tool".⁸ That said, AI may make scams pay better: the blockchain analysts Chainalysis found that crypto scams with links to AI sellers took in about 4.5 times more money per operation, though bigger operations may simply be more likely to buy AI tools.¹¹
• It only covers payment scams. It leaves out card fraud and account takeovers (£703m in 2025), where AI is used to steal details and to fake identities when opening "mule" accounts.⁷ It also leaves out scams nobody reported to their bank.
• It's a snapshot of 2025. The share of spam written by AI roughly tripled in a single year,⁴ so the 2026 figure is likely to be higher.

The fix is simple: measure it. If banks and Report Fraud asked every victim a few specific questions (Did you see a video of a famous person? Did you hear the voice of someone you know? Did you only ever chat in writing?), we would know far more within a year. The Alan Turing Institute has called for a dedicated AI crime taskforce to pull this kind of data together.⁶

7. What this means for you

If around one scam in five involves AI and most victims never notice, then trying to spot the AI isn't a defence. What protects you is the same set of habits that work against every scam, AI or not:

1. Be suspicious of the request, not the presentation. Any message, call, video or advert that leads to moving money, investing or sharing a code deserves a second check, however real it looks or sounds.
2. Check by a different route. Call back on a number you already have, or dial 159, which connects customers of more than 99% of the UK's retail bank current accounts with their bank. For family, use a safe word.
3. Treat celebrity investment adverts as fake and check any investment firm at fca.org.uk.
4. Report it, even if you didn't lose money: Report Fraud on 0300 123 2040 or reportfraud.police.uk, scam texts to 7726 and emails to report@phishing.gov.uk. And if you think AI was involved, say so. That's how the official figures will start to catch up.

For more on the tactics themselves, read how criminals are using AI. And if you run a club or group in East Kent, I give a free talk on AI and the new scams. Ring 07500 511462.

The model is a short Python script, and I'm happy to share it and the full list of assumptions with anyone who'd like to check or improve it: email enquiries@jameshayesconsulting.com.

Sources

1. City of London Police, Report Fraud Annual Assessment FY2025-26, September 2026: 956 AI-flagged reports; 1,240 of 36,464 investment-fraud reports (about 3%) "potentially" AI-enabled; romance-fraud reports mentioning AI rising to 1.5% to 2.0% a month. The 0.17% is my calculation: 956 out of 503,288 fraud and 64,608 cyber crime reports.
2. FBI Internet Crime Complaint Center, 2025 Internet Crime Report, April 2026: 22,364 complaints and $893m in losses with an "AI-related" descriptor, out of 1,008,597 complaints and $20.9bn. Percentages are my calculation.
3. Gallup and the Stop Scams Alliance survey of 5,173 US adults, January to February 2026, reported by NBC News, June 2026.
4. Hao, Tran, Rideout et al., "Do Spammers Dream of Electric Sheep?", ACM Internet Measurement Conference 2025. Data from Barracuda, which employs several of the authors.
5. Lavan, Irvine, Rosi and McGettigan, "Voice clones sound realistic but not (yet) hyperrealistic", PLOS One, 2025.
6. Alan Turing Institute, Centre for Emerging Technology and Security, "AI and Serious Online Crime", March 2025.
7. UK Finance, Annual Fraud Report 2026 (covering 2025), June 2026. Authorised push payment cases and losses by type; unauthorised fraud losses.
8. OpenAI, "Disrupting malicious uses of AI", October 2025.
9. Anthropic, "Detecting and countering misuse of AI: September 2026".
10. Europol, Internet Organised Crime Threat Assessment (IOCTA) 2026, April 2026, p.21.
11. Chainalysis, "Record $17 billion estimated stolen in crypto scams and fraud in 2025", January 2026. Chainalysis sells blockchain analysis tools.