Deepfake Statistics 2026: Fraud and Detection Data

Deepfake statistics in 2026 are dominated by numbers nobody audits: billion-dollar loss figures with no primary source, four-digit surge percentages that contradict each other, and forecasts recycled as measurements. This page does the opposite. Every figure below carries a named source and a publication date, vendor-published numbers are labeled as vendor-published, and the statistics that failed verification are printed in their own table — with the reasons we refused them.

The stakes are not abstract. The World Economic Forum's Global Risks Report 2026 ranks misinformation and disinformation as the world's most severe short-term risk for the second consecutive year, citing AI-generated deepfakes and synthetic audio sophisticated enough to deceive informed audiences. And for the first time in roughly 25 years of reporting, the FBI's Internet Crime Complaint Center broke out AI-enabled fraud as its own category — putting one independently audited dollar figure into a space otherwise filled by vendors marketing detection products.

This guide covers the audited loss data, the vendor-measured attack surface, the peer-reviewed evidence on why humans cannot reliably spot fakes, the regulatory deadlines arriving in the next four weeks, and a verified-versus-recycled trust table you can cite on its own. Fraud defense and detection only — this is a page about protecting organizations, not producing synthetic media.

Start with the one number in this entire space that a government agency audited. The FBI's Internet Crime Complaint Center broke out AI-enabled fraud as a standalone category for the first time in its 2025 Internet Crime Report: 22,364 complaints carrying an AI descriptor, $893.3M in adjusted losses for calendar year 2025. We cross-checked the figure against four independent secondary write-ups of the report; all agree. That works out to roughly $39,900 in adjusted losses per complaint — these are not petty scams.

The category breakdown matters more than the headline. Investment fraud dominates at $632M — roughly 71% of the total. Business email compromise with AI elements accounts for $30M, romance and confidence scams $19M, and AI-involved employment scams roughly $13M. Those four named categories sum to about $694M, leaving roughly $199M spread across other complaint types. Adults 60 and over bore $352M of the total — about 39% of all AI-fraud losses landing on seniors.

The employment-scam line deserves a defensive note. The FBI describes candidates using voice spoofing or deepfake video in online job interviews, and attributes the comparatively low dollar figure to a different criminal objective: network access rather than direct payment. The detection cues IC3 itself publishes for awareness are worth training on — lip-sync mismatches, and physical cues like a cough that doesn't match the video.

Why lead with this figure? Because it is the yardstick that exposes the recycled numbers. When a listicle claims “$1.8 billion lost to voice-cloning scams” or attributes “$2.3 billion in elderly losses” to the FBI, the audited report says otherwise: the FBI's entire AI-fraud total — every category combined — is $893.3M, and its seniors-specific figure is $352M. Both zombie claims fail arithmetic before they fail sourcing. Section 06 handles them in detail.

The attack-surface data comes mostly from vendors who sell detection and verification products — a conflict of interest almost no competing statistics page discloses, and one we label on every figure. With that caveat: Entrust's 2026 Identity Fraud Report (published November 18, 2025, drawing on more than 1 billion identity verifications across 195 countries) found that deepfakes now drive 1 in 5 biometric fraud attempts globally. Deepfaked selfie attempts rose 58% in a year, and injection attacks — feeding fake video directly into verification systems rather than holding a fake up to a camera — rose 40% year over year.

Entrust's document-fraud data shows the same migration from physical to synthetic: digital document forgeries reached 35% of all document fraud in 2025, up from a 29% average across 2022–2024, with physical counterfeits still the larger share at 47% and national ID cards representing 46% of fraudulent document submissions.

On the organizational side, a Gartner survey of 302 cybersecurity leaders across North America, EMEA, and Asia/Pacific (fielded March–May 2025, published September 22, 2025, as reported by Infosecurity Magazine and SC Media) found 62% of organizations experienced a deepfake-enabled attack in the prior 12 months. Within that same sample, 43% reported at least one audio-call deepfake incident and 37% encountered deepfakes in video calls. The voice-and-video-call vector is exactly the pattern behind the Arup Hong Kong case — a $25M wire fraud executed over a deepfaked video call in January 2024 — and the reason deepfake-driven email compromise now gets its own playbook.

Incident-tracking data adds a third lens — and a worked example in why vendor numbers demand scrutiny. Resemble AI's Q3 2025 report (November 2025) recorded 2,031 discrete verified deepfake incidents in that quarter alone, a reported 317% jump versus Q2, with celebrities and public figures making up 48.7% of targets and women targeted 4.5x more often than men in individual-target cases. Five months later, the same vendor's full-year 2025 Threat Report (March 20, 2026) counted 1,567 unique verified incidents for the entire year — 464 fewer than its own single-quarter figure — alongside $1.28B in documented fraud losses and 3,253 total reported incidents including duplicates. Its corporate and consumer fraud sub-categories sum to just 330 incidents, about 21% of the headline count, and more than 80% of tracked incidents carried no disclosed damage figure at all. Both numbers are real, both are correctly sourced — and they cannot both be “unique verified” counts on the same basis. We flag the reconciliation gap instead of quietly picking the cleaner number.

The most-quoted forward-looking figure belongs here too, with its vintage attached: Deloitte's Center for Financial Services projected in May 2024 that US generative-AI-enabled fraud losses could reach $40B by 2027, a 32% compound annual growth rate from a $12.3B 2023 baseline. It is a two-year-old forecast, not a measurement — most competing pages print it as if it were a current result.

The best evidence on human detection is not a vendor study — it is a 2024 peer-reviewed meta-analysis by Diel and colleagues in Computers in Human Behavior Reports, pooling 56 papers and 86,155 total participants. Average human accuracy at telling real from fake: 55.54%, with a 95% confidence interval of 48.87–62.10%. Because that interval crosses 50%, human performance is statistically indistinguishable from a coin flip. Audio is the least-bad medium at 62.08%, video sits at 57.31%, images at 53.16%, and text at 52.00%. Feedback training, AI-assisted support, and caricaturization techniques lifted accuracy to about 65.14% in the same analysis — a gain of under ten percentage points, still far short of reliable.

Vendor testing points the same direction, with the disclosure attached. In iProov's February 11, 2025 study of 2,000 UK and US participants (iProov sells biometric detection; the study was independently covered by multiple press outlets), only 0.1% correctly identified every real and fake item in a mixed image-and-video battery. Participants were 36% less likely to spot a fake in video than in a still image. The sharpest finding is the confidence gap: 60% said they were confident in their ability to detect face swaps — against that 0.1% actual perfect-detection rate. Overconfidence, not ignorance, is the exploitable surface.

The interpretation worth internalizing: detection is not a perception skill your team can meaningfully train into existence. A near-ten-point lift from dedicated training still leaves a one-in-three miss rate, and generation quality keeps improving. That is why every serious defense framework in Section 07 is built on process controls — out-of-band verification, channel integrity — rather than on people “looking harder.” For the adjacent generation-side data (a separate, single-study detection figure lives there — a different study from this 56-paper meta-analysis, and the two should not be conflated or averaged), see our AI video generation statistics roundup.

Four weeks after this post publishes, the EU AI Act's Article 50 transparency obligations become binding — August 2, 2026. From that date, AI-generated or manipulated audio, image, or video content resembling real persons, places, or events must be disclosed by the deployer, even without deceptive intent, with lighter-touch exceptions for evidently artistic, satirical, or fictional works and for authorized law-enforcement use. Penalties run up to €15M or 3% of global annual turnover, whichever is higher, under the Act's penalty framework, per legal analyses (we verified the figure across two independent law-firm write-ups rather than the Official Journal text itself). The European Commission's draft Code of Practice on marking and labeling AI-generated content carries a July 22, 2026 signatory deadline ahead of enforcement — 18 days out as this publishes.

The United States moved earlier on one specific harm. The federal TAKE IT DOWN Act, signed May 19, 2025, required covered platforms to stand up notice-and-takedown processes for nonconsensual intimate imagery — explicitly including AI-generated deepfakes — by May 19, 2026. That deadline has already passed and entered FTC enforcement: platforms must remove reported content, and known identical copies, within 48 hours of a valid report (per the Congressional Research Service's legal sidebar and multiple law-firm analyses).

At the state level, 47 of 50 US states have enacted some form of deepfake-specific legislation as of April 2026, per Ballotpedia and the MultiState tracker — only Alaska, Missouri, and Ohio have not. That is 169 state deepfake laws since 2022, with 82% enacted in just the last two years, and 30 states now carrying election-specific deepfake protections ahead of the 2026 midterms. (One earlier source counted 46 states in spring 2026 — a snapshot-date variance we disclose rather than false-precision away.) For the fuller compliance vocabulary beyond the labeling rule, our EU AI Act compliance glossary covers the adjacent obligations.

The forward-looking read: this is the first period in which the legal exposure from synthetic media is arriving on a fixed calendar while the technical ability to detect it remains statistically near chance. Organizations that publish any AI-generated audio, image, or video with EU exposure have a dated compliance task; organizations defending against fraud get no equivalent deadline dividend — the attacks are already here.

This is the table we could not find anywhere else. The top-ranking “deepfake statistics” pages cite each other in a closed loop, rarely name publication dates, and never disclose which numbers come from vendors with products to sell. Every statistic used in this post is listed below with its named source, its date, and its verification status — so you can cite any row, or check our work.

The pattern the table exposes: the “surge percentage” genre is close to meaningless without three anchors — a baseline period, a measured population, and a disclosed methodology. Sumsub reports 2,100%, Pindrop reports over 1,300%, Entrust reports 40% for injection attacks specifically, and aggregators pass around still-larger four-digit figures nobody can trace (two of them are in the refused table below). These are not measurements of the same thing at different magnitudes; they are measurements of different things, on different traffic, over different windows. Any page that averages them — or cites one as “the” deepfake growth rate — is manufacturing precision that does not exist.

Every statistics page silently omits numbers it doesn't trust. We think the omissions are the story. The claims below circulate widely across 2026 “deepfake statistics” pages — several rank on the first page of search results — and each one failed verification in a specific, checkable way. If you have cited one of these, this is the row to bookmark.

Two disclosures for completeness. First, the Deloitte $40B figure appears in both tables deliberately: as a dated May 2024 forecast it is usable; as a claimed 2026 measurement it is refused. Second, the Resemble AI quarterly-versus-annual inconsistency (Section 02) is the clearest illustration of why this section exists — when a single vendor's own Q3 count exceeds its own full-year “unique verified” count, the lesson is not that the vendor is lying, but that incident-tracking methodology is young, unstandardized, and unaudited. Treat every non-government number in this space as directional.

The through-line of every verified number on this page: humans cannot be the detection layer, so process has to be. Four moves follow directly from the data — none of them require buying a detection product first.

Fraud defense is also a measurement problem. The organizations that catch synthetic-identity and payment fraud early are the ones instrumenting their funnels well enough to see anomalies — the same discipline covered in our payment fraud and chargeback prevention playbook and in the measurement foundations our analytics engagements build. And if the Article 50 deadline or verification-first process design lands on your desk, that is the kind of operational AI-governance work our AI transformation engagements are scoped for — controls and workflows first, tooling second.