Deepfake Voices and Faces in Fraud: What Is Known in 2026

The short version

Fraud that uses synthetic voices and video has moved from rare to routine in less than a decade. Most of it is still conventional social engineering: a believable story, urgency, and a request to move money. A cloned voice or a faked face makes that request more convincing. Losses that law enforcement can confirm are large but still a small share of all fraud. Many of the alarming growth figures in circulation come from companies that sell detection and identity-verification products.

How the technology works

A voice clone is a model trained to copy one person's voice. Modern systems need only seconds to minutes of clean audio, which is often available from social media, podcasts, earnings calls or voicemail greetings. The FBI's December 2024 public service announcement (I-120324-PSA) says criminals "generate short audio clips containing a loved one's voice to impersonate a close relative in a crisis situation."[1] Text-to-speech clones read out whatever the scammer types. Speech-to-speech converters change a live speaker's voice into the target's voice in real time.

Video deepfakes work in a similar way. Face-swap software maps a target's face onto an actor's head, frame by frame. Some tools run fast enough for live video calls. Others produce pre-recorded clips that are played into a meeting or posted as an advertisement. Lip movement, lighting at the edges of the face, and sudden head turns are where these tools still tend to fail, although the failures are becoming harder to spot.

The tools are cheap. A 2025 ICAEW interview with cyber security professor Oli Buckley noted that paid voice-cloning apps cost as little as US$5 to US$10 a month, and many platforms are free.[2] UNODC's October 2024 report on Southeast Asian organised crime identified more than ten deepfake software vendors specifically targeting criminal groups in that region. It also recorded a 600% increase in deepfake-related mentions in monitored criminal Telegram channels and underground forums between February and June 2024.[3][4]

How criminals use it

Executive impersonation and business email compromise. The best-known case is Arup. In January 2024, a finance employee in the engineering firm's Hong Kong office made 15 transfers totalling HK$200 million (about US$25.6 million) to five local bank accounts.[5][6] He did so after a video conference in which the company's UK-based chief financial officer and other colleagues appeared. Hong Kong police said every other participant was fake.[6][7] Acting Senior Superintendent Baron Chan told reporters the scammers had found "publicly available video and audio of the impersonation targets via YouTube" and that the videos were pre-recorded and did not involve real interaction with the victim.[8] The employee had first suspected a phishing email; the call removed his doubts.[6] Arup confirmed it was the victim in May 2024.[9] Police reported no arrests at the time.[8]

Earlier cases set the pattern. In 2019 the Wall Street Journal reported that fraudsters used a cloned voice of a German parent company's chief executive to get the head of a UK energy subsidiary to transfer about €220,000.[2] In early 2020 a cloned company director's voice and emails from a fictitious lawyer were used to move about US$35 million (see The UAE's US$35 Million Deepfake Voice Fraud).[10] The FBI's 2025 Internet Crime Report says chat generators can produce "official-sounding emails mimicking a company's CEO" and that voice cloning is used to request wire payments. It also records more than US$30 million in business email compromise losses in 2025 with a reported AI link.[11] That is about 1% of the US$3.05 billion total BEC losses for the year.

"Family emergency" and grandparent scams. A cloned voice of a grandchild, child or friend claims to be in jail, in hospital or kidnapped, and asks for money quickly. The FBI's 2025 report records more than US$5 million in losses to these "distress" scams.[12] It notes they are "evolving to mimic other family members or close friends in different types of emergency scenarios." In December 2025 the FBI also warned that criminals were using altered "proof-of-life" photos and videos in virtual kidnapping ransom schemes.[11]

Romance scams. Fake dating profiles use AI-generated photos. Conversations are drafted or translated by chatbots. Face-swapping lets a scammer appear on video as the person in the profile. The FBI reported more than US$19 million in 2025 romance and confidence losses with a likely AI link.[11] That small figure almost certainly undercounts the problem, because victims rarely know AI was involved.

Investment scams using deepfaked public figures. Fake videos of celebrities, chief executives and financial commentators promoting trading platforms or "investment clubs" circulate as social media ads. The FBI says these clubs "employ AI-generated videos and voices of celebrities, CEOs, or trusted figures." It recorded more than US$632 million in investment-fraud losses with a reported AI link in 2025, the largest AI-linked category by far. The Bureau noted that overall investment losses exceeded US$8 billion, "demonstrating that many victims do not realize the extent AI may be involved."[11]

Impersonation of officials. On 15 May 2025 the FBI warned (I-051525-PSA) that since April 2025, "malicious actors have impersonated senior US officials" using text messages and "AI-generated voice messages" to build rapport and then take over victims' accounts. Many targets were current or former federal and state officials and their contacts.[13][14] The FBI issued a follow-up warning on 19 December 2025 saying the campaign was continuing.[11]

Sextortion. Criminals use face-swapping and "nudify" tools to create explicit images from ordinary photos, then demand payment. Sometimes they use synthetic images to make a fake persona believable. The FBI received more than 75,000 sextortion submissions in 2025.[11] A UNODC technical brief from September 2025 noted "indications of scam centres in Southeast Asia exploiting AI to conduct sextortion."[15]

Bypassing identity checks (KYC). Banks, crypto exchanges and fintech apps ask new customers to upload an ID and take a selfie or short video. Fraudsters answer with AI-generated documents, synthetic faces, or "injection attacks" that feed a deepfake video directly into the verification software instead of the phone's camera. In November 2024 the US Financial Crimes Enforcement Network issued alert FIN-2024-Alert004.[16] It said it had seen an increase since 2023 in suspicious activity reports describing deepfake media, "particularly the use of fraudulent identity documents to circumvent identity verification."[17][18] FinCEN asked banks to tag such reports "FIN-2024-DEEPFAKEFRAUD."[19] The FBI's 2025 report also describes job interviews in which "the actions and lip movement of the person interviewed on-camera do not completely coordinate with the audio," a tactic linked to remote-worker fraud, including schemes by North Korean IT workers.[11]

Scale: what the numbers do and do not show

The most reliable figures come from law enforcement and regulators, and they are partial.

FBI IC3, 2025: 22,364 complaints containing a reference to AI, with US$893.3 million in adjusted losses.[11][20][21] This was the first year the descriptor appeared in the annual report.[22] It captures only cases where the complainant mentioned AI (see What FBI Data Says About Fraud, Scams and AI).
FinCEN: confirms a rise in deepfake-related suspicious activity reports since 2023, but has published no counts.[23]
UNODC: reports the 600% rise in deepfake-related mentions in criminal channels in early 2024. Its October 2024 report also cites a 1,530% increase in deepfake-related crime in the Asia-Pacific region between 2022 and 2023, drawn from industry data.[24][25]

Industry figures are larger and less certain. Each comes from a company that sells detection, identity-verification or security products. Each describes only what that company's own customers saw.

Entrust's 2026 Identity Fraud Report, based on more than one billion identity verifications, found deepfakes in one in five biometric fraud attempts. It reported deepfake selfies up 58% and injection attacks up 40%. This measures attempts inside Entrust's network, not fraud across the economy.[26]
Shufti's Identity Fraud Index projects a 495% rise in deepfake identity fraud in 2026 and a 3,892% rise in document deepfakes. These are annualised projections from five months of its own data, not observed annual results.[27][28][29]
Pindrop's 2025 Voice Intelligence & Security Report, released in June 2025 and based on more than 1.2 billion customer calls, said deepfake fraud attempts against its call-centre customers rose more than 1,300% in 2024, from about one a month to seven a day.[23]
Signicat's 2024 report The Battle Against AI-Driven Identity Fraud, a survey of more than 1,200 fraud decision-makers in the EU and UK, found deepfakes rose from 0.1% to 6.5% of fraud attempts over three years, a 2,137% increase. This figure is widely repeated without attribution.
Resemble AI's H1 2026 Deepfake Threat Report verified 821 attacks from 1,760 news reports in six months.[26]
Surfshark, using public incident databases, has put documented global deepfake-fraud losses at US$2.19 billion (an earlier edition) and at least US$3.7 billion from January 2020 to June 2026 (a later edition). It says most of those losses were recorded from 2025 onwards.[30][31] The two editions disagree because they cover different periods. Both count only publicly reported losses.

These sources point the same way: more attempts, more losses and better fakes. Their percentages should not be compared or added together. Very large growth rates often reflect small starting numbers, changes in what a vendor's tools can detect, or growth in the vendor's customer base. Some widely shared claims, such as "fewer than 5% of voice-clone victims report,"[20] circulate without a traceable primary source and should be treated as unverified.

Detection and defence

Detection software exists for synthetic voice, face-swaps and injection attacks, and banks increasingly require it. It is not reliable on its own. Detectors trained on one generation of tools miss the next. In an iProov study released in February 2025, only 0.1% of 2,000 UK and US consumers, who had been told to look for deepfakes, correctly identified every real and fake item shown to them, so human judgement is not a dependable backstop either.

The defences that work best are procedural:

Verify through a separate channel. Call the person back on a number already on file, not one given in the message. In the Arup case the video call itself was the fake verification step.[32]
Require two people to approve large or unusual payments, and do not allow exceptions for "confidential" deals or urgent requests from senior executives.
Use family code words. The FBI recommends agreeing a secret word or phrase with relatives for emergencies.[12]
Limit public audio and video where practical, and make social media accounts private. The FBI recommends both.[33]
Strengthen verification: liveness checks that resist injection attacks, phishing-resistant multi-factor authentication, and treating voice authentication as weak. FinCEN lists live verification checks and phishing-resistant MFA among its recommended practices.[34]
Report quickly. The FBI's Recovery Asset Team froze US$679 million of US$1.16 billion in attempted thefts in 2025, a 58% success rate.[11][35] Its success depends on how fast a report reaches it.

Regulatory responses

In February 2024 the US Federal Communications Commission ruled that AI-generated voices in robocalls count as "artificial" under the Telephone Consumer Protection Act, which makes unsolicited AI-voice robocalls illegal without consent. The TAKE IT DOWN Act, signed in May 2025, makes it a federal crime to publish non-consensual intimate images, including AI-generated ones, and requires platforms to remove them on request. FinCEN's alert set reporting expectations for banks but created no new legal duty.[18] In the European Union, the AI Act's transparency rules require deepfake content to be labelled; they apply from August 2026. Elsewhere, responses have mostly come through existing fraud, cybercrime and data-protection laws, public warnings, and banking rules that move away from SMS one-time codes and voice verification. The UAE's measures are covered in The UAE's US$35 Million Deepfake Voice Fraud.

What is established and what is not

It is established that voice cloning and video deepfakes have been used in confirmed frauds, some worth tens of millions of dollars. The tools are cheap and widely available. Law enforcement in the US, Hong Kong, the UAE and elsewhere has documented specific cases. It is an estimate that deepfake fraud is growing by hundreds or thousands of percent a year; that rests mostly on vendor data. It is uncertain what share of total fraud losses AI actually accounts for, because almost no reporting system can tell whether a scammer used AI unless the victim noticed.