Deepfake fraud, by the numbers.

Deepfake statistics travel fast and lose their sources on the way. Most of the numbers quoted in sales decks trace back to a handful of reports, each with its own method and its own blind spots. This note collects the figures we are willing to cite, names where each one comes from, and flags what they can and cannot tell you.

How much is being lost

$12.3B to $40B

US fraud losses enabled by generative AI, 2023 projected to 2027Deloitte Center for Financial Services, May 2024

32%

Compound annual growth rate implied by that projectionDeloitte Center for Financial Services, May 2024

$200M+

Losses attributed to deepfake incidents in the first quarter of 2025 aloneResemble AI, Q1 2025 Deepfake Security Report

$450,000

Average loss per affected organization; one in ten reported losing more than $1MRegula, Deepfake Trends 2024

Read the Deloitte figure as a scenario, not a measurement. It models how fraud losses could grow if generative tools keep lowering the cost of a convincing attack. The Regula figure is a survey of businesses that had already been hit, so it describes the cost of a successful attack rather than the average exposure of a company.

How often it happens

Every 5 minutes

Rate of deepfake attempts observed across identity verification traffic in 2024Entrust, 2025 Identity Fraud Report

244%

Year-over-year increase in digital document forgeries in 2024Entrust, 2025 Identity Fraud Report

10x

Growth in deepfake incidents detected globally from 2022 to 2023Sumsub, Identity Fraud Report 2023

+1,740%

Growth in North America over the same period; Asia-Pacific grew 1,530% and Europe 780%Sumsub, Identity Fraud Report 2023

These come from verification vendors measuring their own traffic. They are real observations, but of a specific slice: people trying to pass an identity check. Growth rates that large mostly reflect how small the base was in 2022. The useful signal is the direction and the fact that the same pattern shows up in independent data sets.

Who it touches

1 in 4

Adults who had experienced an AI voice scam themselves or through someone they know, across 7,054 respondents in seven countriesMcAfee, The Artificial Imposter, May 2023

77%

Share of those targeted who lost moneyMcAfee, The Artificial Imposter, May 2023

3 seconds

Audio needed to produce a usable voice clone in McAfee's own testMcAfee, The Artificial Imposter, May 2023

How well people spot it

55.5%

Pooled human accuracy at telling deepfakes from genuine media across 56 studies; not significantly better than chanceDiel et al., Computers in Human Behavior Reports, 2024

57%

Pooled human accuracy for video specificallyDiel et al., Computers in Human Behavior Reports, 2024

This is the number that matters most for anyone designing a review process. A trained reviewer looking at a single clip is close to a coin flip, and the studies behind this figure used material that is now several generations old. Manual review is still essential, but it needs evidence in front of it: provenance, metadata, spectral and temporal signals, and a record of what was checked. That is the premise the Plurall AI platform is built on.

Numbers we chose not to repeat

Several figures circulate without a traceable origin, including claims that humans catch high-quality deepfakes about a quarter of the time, that attacks grow 900% a year, or that a fixed number of companies are targeted each day. We could not find primary sources for them and have left them out. If a statistic on this page cannot be followed back to a report with a method, tell us and we will remove it.

Figures are reproduced as published by their sources, with the publication date noted. They are not Plurall AI measurements. No proprietary model internals or customer-specific details are included.