Deepfake Statistics 2026: Verified Figures on Fraud, Losses and Detection

How These Deepfake Statistics Were Selected

Deepfake numbers circulate without dates, sources or definitions, and many are vendor marketing repeated until they look like fact. Every figure on this page was checked against its primary source on 19 October 2026: a government report, a regulator's release, a peer-reviewed or arXiv paper, or the survey owner's own publication. Each entry says who measured it, when, and what was actually counted.

Read the categories differently. Complaint data such as the FBI's undercounts, because most victims never report. Survey data reflects what respondents say. Projections are models, not measurements. Benchmark results describe a specific test set. We left out widely repeated figures we could not open in the original source.

How Much Money Is Lost to Deepfake Fraud?

The best measured figure is the FBI's. Its 2025 Internet Crime Report, published in April 2026, recorded 22,364 complaints that referenced artificial intelligence, with adjusted losses exceeding $893 million. Within that total, investment complaints with an AI nexus passed $632 million, business email compromise involving AI exceeded $30 million, confidence and romance scams with a likely AI nexus exceeded $19 million, and so-called distress scams using cloned voices of relatives exceeded $5 million. The FBI notes that overall investment fraud losses topped $8 billion, which suggests many victims never realised AI was involved.

For scale, the same report counted 1,008,597 complaints and $20.877 billion in total losses in 2025, a 26% rise on 2024. The FTC's Consumer Sentinel data, presented to Congress on 25 March 2026, showed $15.9 billion in reported fraud losses for 2025, including more than $3.5 billion to imposter scams, the most reported fraud type. Neither total is deepfake-specific, but impersonation is where cloned voices and faces are used.

Single incidents can be large. In early 2024 an employee in the Hong Kong office of engineering firm Arup transferred HK$200 million, about US$25 million, in 15 transactions after a video call in which, according to Hong Kong police, every other participant was fake.

Projections are higher and should be labelled as such. Deloitte's Center for Financial Services projected in May 2024 that generative AI could push US fraud losses from $12.3 billion in 2023 to $40 billion by 2027, a 32% compound annual growth rate.

How Common Are Deepfake Attacks?

Volume estimates are rising steeply. The UK government said on 5 February 2026 that an estimated 8 million deepfakes were shared in 2025, up from 500,000 in 2023, and announced a deepfake detection evaluation framework with Microsoft, academics and other experts after a four-day detection challenge with more than 350 participants. The figure is an estimate rather than a count.

Regulators see the same trend in financial crime. FinCEN's alert FIN-2024-Alert004 of 13 November 2024 reported an increase, beginning in 2023 and continuing in 2024, in suspicious activity reports describing deepfake media in fraud, often altered or synthetic identity documents and selfies used to get past verification. It asked banks to tag such reports with the key term FIN-2024-DEEPFAKEFRAUD and listed red flags including third-party webcam plugins during live checks and photos flagged by deepfake detection software.

What Percentage of Fraud Attempts Involve Deepfakes?

About 6.5%, or roughly 1 in 15, in one European financial-sector survey. Signicat's February 2025 release, based on a 2024 Censuswide survey of 1,206 fraud decision-makers at banks, insurers, payment providers and fintechs in seven European countries, said deepfakes had risen from 0.1% of fraud attempts three years earlier. Treat this as respondents' estimates, not transaction data. The release also quotes 2,137% growth over three years, a figure that does not follow arithmetically from the two shares it gives, so we do not repeat it as a measure.

Shares vary heavily by channel. Onboarding and account recovery see deepfake selfies and injected video; contact centres see cloned voices; finance teams see impersonated executives. A single percentage across all fraud hides where the risk concentrates for a given business.

Can People Detect Deepfakes? The Numbers

Mostly not. In a study published on 12 February 2025, iProov tested 2,000 consumers in the UK and US and found that only 0.1% correctly identified every real and fake stimulus, even though participants were told to look for deepfakes. Participants were 36% less likely to identify a synthetic video correctly than a synthetic image, more than 60% were confident in their answers regardless of accuracy, and 22% had never heard of deepfakes before the study.

NIST AI 100-4 (November 2024) summarises academic work the same way: a 2024 study across audio, visual and audiovisual content found human performance near chance. Training helps people notice context and process red flags. It does not turn them into reliable pixel-level detectors.

How Accurate Is Deepfake Detection Software?

It depends on the detector and the test, and the spread is wide. Deepfake-Eval-2024, an in-the-wild benchmark of 45 hours of video, 56.5 hours of audio and 1,975 images collected from 88 websites in 52 languages, found that open-source detectors lost about 50% of their AUC on video, 48% on audio and 45% on images compared with earlier academic benchmarks. Commercial and fine-tuned models did better but, the authors write, do not yet match deepfake forensic analysts. NIST AI 100-4 cites cross-generator image detector accuracy of 61% to 70% in a 2023 paper.

Commercial systems differ sharply too. On the public VendorBench-100 benchmark (v2, 29 July 2026), which tested 36 systems on one adversarial set of 100 still images and ranked them by Matthews correlation coefficient, Neural Defend scored MCC 0.876 with ROC-AUC 0.979, Reality Defender 0.646, Gemini 0.389, Hive 0.290, Sightengine 0.258 and OpenAI's GPT 0.135. The paper warns accuracy and F1 are unreliable on that imbalanced set. One VendorBench author also co-wrote Neural Defend's ARCAS 1B paper; the corpus and results are public so anyone can rerun them, and the benchmark does not test video, audio or injection attacks.

How Many Deepfake Laws Exist?

In the US, Ballotpedia reported on 30 July 2026 that 49 states had passed at least one deepfake law since 2019, 48 had laws on sexual deepfakes and 33 on political deepfakes, with 58 deepfake bills enacted in 2026 alone. At federal level, the TAKE IT DOWN Act was signed on 19 May 2025. Outside the US, China's AI content labelling measures took effect on 1 September 2025, India's synthetic content rules on 20 February 2026, and EU AI Act Article 50 on 2 August 2026. The country-by-country deepfake law guide on this site has the detail.

What These Statistics Mean for Fraud and Identity Teams

Three conclusions hold across the sources. Reported losses are a floor, because the FBI's own commentary suggests many victims never identify the AI element. People are a weak control, so processes cannot depend on staff noticing a fake face or voice. And detector quality varies by an order of magnitude, so vendor claims need testing on your own media with metrics that respect class imbalance.

Neural Defend's NeuroVerify is built for the channels where these losses start. One API call checks a face image, video or audio with six signals, including Live Presence Verification, Face Manipulation Detection and AI Generative Model Analysis, and returns a unified risk score with signal-level evidence in under 200 ms p95 on eligible image requests. It integrates through mobile SDKs for Android, iOS and Flutter, backend SDKs for Python, TypeScript, Go and Java, or REST, and deploys in the cloud, a private VPC, on premises or air-gapped.