Deepfake fraud figures circulate widely, but many cannot be traced to the body that collected them, and the large headline numbers often mix forecasts, vendor telemetry and reported losses as if they were the same thing. This page gathers deepfake and AI-enabled fraud statistics only from primary sources: the FBI Internet Crime Complaint Center (IC3), the US Financial Crimes Enforcement Network (FinCEN), the UK communications regulator Ofcom, peer-reviewed research, and the original reports of identity verification and voice security vendors, each labelled as vendor data. Every figure carries its publisher, report name and as-of period. It is written for executives sizing the risk, fraud and security teams building a business case, and analysts who need numbers that survive scrutiny. For how the attacks work and which controls stop them, see the Deepfake and AI Impersonation Guide. Last reviewed: November 2026.
Key takeaways
- Reported AI-linked losses are already near $900 million a year in the US. The FBI IC3 Internet Crime Report 2025 recorded 22,364 complaints referencing AI, with $893,346,472 in losses.
- Investment scams dominate the money. AI-linked investment fraud accounted for $632,041,188 of those losses in the IC3 2025 data, about 71% (calculated).
- Deepfakes are now routine in identity verification attacks. Entrust's 2026 Identity Fraud Report found deepfakes in 1 in 5 biometric fraud attempts between September 2024 and September 2025.
- People cannot reliably spot them. In an iProov study of 2,000 UK and US consumers, only 0.1% correctly classified every real and fake image and video, and a peer-reviewed study found listeners identified deepfake speech 73% of the time.
- Forecasts are not losses. The widely quoted $40 billion figure is Deloitte's 2024 projection of generative AI-enabled US fraud losses by 2027, not a measured total.
Headline statistics
| Statistic | Figure | Source, publisher, period |
|---|---|---|
| Complaints referencing AI reported to the FBI | 22,364 | Internet Crime Report 2025, FBI IC3, calendar year 2025 |
| Losses in complaints referencing AI | $893,346,472 | Internet Crime Report 2025, FBI IC3, 2025 |
| Share of biometric fraud attempts that were deepfakes | 1 in 5 | 2026 Identity Fraud Report, Entrust (vendor data), September 2024 to September 2025 |
| Contact centre calls that were synthetic or replayed voice, end of 2024 | 0.94% (1 in 106) | Voice Intelligence and Security Report, Pindrop (vendor data), 2024 |
| UK adults who say they saw a deepfake in the previous six months | 43% | Deepfakes research, Ofcom, survey June 2024 |
| Consumers who correctly classified every real and fake stimulus | 0.1% | Deepfake detection study, iProov (vendor-commissioned), published February 2025 |
| Forecast generative AI-enabled US fraud losses by 2027 | US$40 billion (from US$12.3 billion in 2023) | Center for Financial Services, Deloitte, published May 2024 (forecast) |
Reported losses
The FBI IC3 report is the best recurring source of reported losses. Its "AI Related" descriptor marks complaints whose information "contains a reference to artificial intelligence", so it captures AI involvement as victims describe it, not forensic confirmation of a deepfake. It undercounts cases where victims did not realise AI was used.
| Crime type (AI-related complaints) | Complaints | Losses |
|---|---|---|
| Investment | 4,356 | $632,041,188 |
| Business email compromise | 135 | $30,256,592 |
| Tech and customer support | 574 | $19,457,078 |
| Confidence and romance | 626 | $19,041,653 |
| Employment | 691 | $12,550,185 |
| Government impersonation | 260 | $7,061,628 |
| Extortion | 1,764 | $2,940,642 |
Source for the table: FBI IC3 Internet Crime Report 2025, calendar year 2025.
- Few cases, large losses in BEC: only 135 AI-related BEC complaints produced over $30 million in losses, an average of roughly $224,000 per complaint (calculated from IC3 2025). The IC3 report notes voice cloning can be used to request wire payments. See the Business Email Compromise and Fraud Statistics page for BEC overall.
- Extortion is high volume, low value: 1,764 AI-related extortion complaints, the second highest count, but under $3 million in losses (IC3 2025).
- Older victims carry a large share: people aged 60 and over filed 3,143 AI-related complaints with $352,496,231 in losses in IC3 2025, about 39% of all AI-related losses (calculated). IC3 also reports over $5 million in claimed losses to "distress" scams, where cloned voices mimic a relative in trouble.
- Employment scams target access, not just money: IC3 2025 describes deepfakes and voice spoofing in online job interviews, where the aim appears to be access to private networks. See the Hiring Fraud and Candidate Identity Verification Guide.
Single incidents
Individual cases can exceed whole categories in the IC3 data. In the Arup deepfake fraud of 2024, a Hong Kong finance worker joined a video call with deepfaked colleagues, including the CFO, and made 15 transfers totalling HK$200 million (about US$25.6 million) before checking with head office. In May 2025 the FBI warned in PSA I-051525-PSA that, since April 2025, actors had been sending AI-generated voice messages impersonating senior US officials to steal account credentials.
Forecast losses
Deloitte's Center for Financial Services published, on 29 May 2024, a prediction that generative AI could enable fraud losses of US$40 billion in the United States by 2027, up from US$12.3 billion in 2023, a compound annual growth rate it gives as 32%. Treat this as a scenario for planning, and say "forecast" whenever you quote it.
Attack volumes in identity verification
Identity verification vendors see attempted attacks against onboarding and authentication, including those that fail. Their data shows volume and technique trends; it is not a measure of losses and reflects each vendor's customer base.
| Statistic | Figure | Source, publisher, period |
|---|---|---|
| Deepfakes as a share of biometric fraud attempts | 1 in 5 | 2026 Identity Fraud Report, Entrust, September 2024 to September 2025 |
| Growth in deepfaked selfies | Up 58% in 2025 | 2026 Identity Fraud Report, Entrust |
| Growth in injection attacks | Up 40% year on year | 2026 Identity Fraud Report, Entrust |
| Digital forgeries as a share of document fraud | 35% in 2025, up from a 29% average for 2022 to 2024 | 2026 Identity Fraud Report, Entrust |
| Growth in UK deepfake fraud attempts | Up 94% in 2025 | Identity Fraud Report 2025-2026 release, Sumsub, 25 November 2025 |
| Growth in multi-step "sophisticated" fraud globally | Up 180% in 2025 | Identity Fraud Report 2025-2026 release, Sumsub |
The rise in injection attacks matters more than the deepfake share alone: injection bypasses the camera by feeding synthetic video straight into the capture flow, so presentation attack detection by itself does not stop it. FinCEN alert FIN-2024-Alert004 (13 November 2024) lists use of a virtual webcam during live verification as a red flag and reports increased suspicious activity reporting describing deepfake media from 2023 into 2024. US financial institutions are asked to use the key term "FIN-2024-DEEPFAKEFRAUD" in those reports. See the Deepfake-Resistant Identity Verification Guide.
Voice cloning and contact centres
Voice deepfakes target phone banking, insurance claims and help desks. The most detailed published series comes from one voice security vendor's call telemetry.
| Statistic | Figure | Source, publisher, period |
|---|---|---|
| Non-live calls (synthetic or replayed voice), end of 2024 | 0.94% of calls, 1 in 106; 6.8 times the start of the year | Voice Intelligence and Security Report, Pindrop, 2024 (over 130 million calls analysed) |
| Deepfake activity in Q4 2024 | Passed 2% of calls, doubling within the quarter | Voice Intelligence and Security Report, Pindrop |
| Deepfake attacks per day at large national banks | 5.5 in Q4 2024, up from 1.6 at the start of 2024 | Voice Intelligence and Security Report, Pindrop |
| Growth in synthetic voice fraud | Up 149% at banks and 475% at insurers in 2024 | Voice Intelligence and Security Report, Pindrop |
| Forecast growth in deepfake fraud | Up 162% in 2025 (vendor forecast) | Voice Intelligence and Security Report, Pindrop |
These figures come from the vendor's own customers, mainly US financial services, and should not be generalised to all organisations. They do show why help desk password and MFA resets based on a voice call are a weak point. See the Account Recovery and Help Desk Security Guide.
Detection rates: people and exposure
| Statistic | Figure | Source, publisher, period |
|---|---|---|
| Participants who correctly classified all real and fake images and videos | 0.1% of 2,000 | iProov study, UK and US, published 12 February 2025 |
| Relative difficulty of video | 36% less likely to identify a synthetic video than a synthetic image | iProov study, 2025 |
| Confidence regardless of accuracy | Over 60% confident in their detection skills | iProov study, 2025 |
| Accuracy identifying deepfake speech | 73% (529 participants, English and Mandarin); training with examples gave only a small improvement | Mai et al., PLOS ONE, 2023 |
| UK adults confident they could spot a deepfake | 9% | Ofcom, survey June 2024 |
| UK adults who say they saw a deepfake in the last six months | 43% (children aged 8 to 15: 50%) | Ofcom, survey June 2024 |
| Children aged 8 to 15 who encountered a deepfake scam advert | 32% | Ofcom, survey June 2024 |
The pattern across independent and vendor studies is consistent: people overestimate their ability, video is harder than still images, and training improves detection only slightly. This is the strongest statistical argument for process controls (call-back verification, dual approval) over "spot the fake" awareness alone. See the Spotting Deepfake Scams Guide for what training can realistically achieve.
Methodology notes and caveats
- Reported versus attempted: IC3 counts complaints victims file, with losses as reported. Vendor figures count attempts their systems saw, most of which failed. Do not add them together.
- AI is self-described in IC3: the AI descriptor depends on what victims wrote, so it is a floor, not a full count of deepfake fraud.
- Vendor base effects: Entrust, Pindrop and Sumsub figures describe their own customers and markets. Growth percentages from a small base can look dramatic.
- Forecasts: Deloitte's $40 billion is a projection of generative AI-enabled fraud broadly, not deepfakes alone. The page's 32% growth rate also differs slightly from what its own endpoints imply, so quote the endpoints and label them as a forecast.
- Commissioned studies: the iProov detection study was commissioned by a vendor of biometric verification; its finding is consistent with the peer-reviewed speech study and Ofcom's confidence data.
- Refresh cadence: IC3, Entrust, Pindrop and Sumsub publish annually, so this page can be updated each year with like-for-like series.
What the numbers mean for defenders
| What the data shows | Control it supports | Framework anchor |
|---|---|---|
| Over $30 million in AI-related BEC losses from 135 complaints (IC3 2025); Arup's HK$200 million | Call-back to a known number and dual approval for payments and bank detail changes, regardless of who appears on a call | NIST CSF 2.0 PR.AT-01; NIST SP 800-53 AT-2(3) |
| Deepfakes in 1 in 5 biometric fraud attempts; injection attacks up 40% (Entrust) | Injection attack detection and server-side liveness in onboarding, independently tested | NIST SP 800-63A; NIST SP 800-53 IA-12 |
| Synthetic voice at over 2% of calls in Q4 2024 for one vendor's customers | No password or MFA reset on voice alone; out-of-band verification for help desk requests | NIST SP 800-53 IA-5; MITRE ATT&CK T1566.004 Spearphishing Voice |
| 0.1% of people classified every stimulus correctly; 73% speech accuracy | Process controls over detection by eye; training focused on procedures, not spotting | NIST SP 800-53 AT-2(3) |
| Deepfaked job interviews aimed at network access (IC3 2025) | Identity verification at hiring and at first device enrolment | NIST SP 800-53 PS-3 |
For a response plan when a deepfaked executive request gets through, see the Deepfake CEO Fraud Response Playbook. To evaluate detection products, see the Deepfake Detection Buyer's Guide.
Practitioner checklist
- Quote every deepfake figure with publisher, report and period, and label vendor data and forecasts as such.
- Use IC3 AI-related losses for US reported loss, vendor telemetry for attack trends, and never combine the two.
- Require call-back verification and dual approval for payment and bank detail changes, regardless of the channel the request arrived on.
- Remove voice-only identity checks from help desk credential and MFA resets.
- Ask identity verification providers for injection attack detection evidence, not just liveness claims.
- Build FinCEN's red flags (virtual webcams, inconsistent documents, refused MFA) into onboarding review if you are a US financial institution.
- Train staff on procedures to follow when a request is unusual, since detection by eye is unreliable.
- Refresh this page's figures each year when IC3, Entrust, Pindrop and Sumsub publish.
Standards and references
- FBI IC3: Internet Crime Report 2025
- FinCEN: Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (FIN-2024-Alert004)
- FBI IC3: Senior US Officials Impersonated in Malicious Messaging Campaign (I-051525-PSA)
- Ofcom: Deepfakes: demean, defraud, disinform
- PLOS ONE: Warning: Humans cannot reliably detect speech deepfakes (Mai et al.)
- Deloitte: Generative AI is expected to magnify the risk of deepfakes and other fraud in banking
- Entrust: 2026 Identity Fraud Report announcement
- Pindrop: Voice Intelligence and Security Report 2025
Further sources used: iProov: deepfake detection study (February 2025) · Sumsub: Identity Fraud Report 2025-2026 release.
Related NHI Mgmt Group resources: Deepfake and AI Impersonation Guide · Executive Deepfake and CEO Fraud Guide · Business Email Compromise and Fraud Statistics · Shadow AI Statistics