Deepfakes stopped being a novelty act somewhere around 2023 and became an industry: a fraud industry, an abuse industry, and, on the other side, a detection industry racing to keep up. The numbers below chart all three, with the projections labeled as projections.
The short answer: roughly 8 million deepfakes circulate online, up from about 500,000 in 2023, growth of nearly 900% a year. A deepfake fraud attempt now occurs about every five minutes, humans correctly identify high-quality deepfake video only 24.5% of the time, and 96 to 98% of deepfake videos are non-consensual intimate content, with roughly 99% of those victims women.
As always, I sorted the numbers the way a data researcher has to: audited and study-backed figures first, projections flagged, and every stat naming who measured it. Sources at the bottom.
Table of contents · 8 sections
Deepfakes in Numbers: Editor's Choice
- →Deepfake files online grew from ~500,000 in 2023 to a projected 8 million in 2025, roughly 900% annual growth. (DeepMedia; Europol)
- →A deepfake fraud attempt occurs about every 5 minutes. (Entrust)
- →Only 0.1% of people can reliably spot every deepfake, while over 60% are confident they can. (iProov)
- →Human accuracy on high-quality deepfake video: 24.5%. (iProov)
- →One deepfake video call cost engineering firm Arup $25 million. (Financial Times)
- →62% of organizations experienced a deepfake incident in the past year; the average identity-fraud attempt costs businesses about $450,000–500,000. (Gartner; Regula)
- →96–98% of deepfake videos are non-consensual intimate content, and ~99% of victims are women. (Sensity; Home Security Heroes)
How many deepfakes are there?
Nobody can count them all, but the best-tracked estimates agree on the shape of the curve: vertical.
1. Deepfakes online grew from ~500,000 in 2023 to a projected 8 million in 2025.
The most-cited volume estimate, and notably one that serious institutions repeat: Europol flags it, and so does the UK government. It tracks with DeepMedia's observed doubling of deepfake volume roughly every six months.
2. That's a growth rate of roughly 900% per year.
For scale, almost no cyber threat category grows at triple digits, and this one grows at nearly four. At this pace, today's 8 million becomes tens of millions within a year, which is why every fixed count on this page has an expiry date.
3. Deepfake videos specifically increased about 550% between 2019 and 2024.
Around 95,800 deepfake videos were counted online in 2023, up from under 15,000 in 2019. Video was the original format; audio has since become the faster-growing attack channel, more on that below.
4. The tools behind the curve cost as little as $50 per campaign, and many are free.
The growth isn't mysterious: the barrier to entry collapsed. Voice cloning, face swapping, and video generation are now consumer products. What used to require a VFX studio requires a browser tab.
Can people actually spot deepfakes?
This is the gap that makes every other number on this page dangerous: the distance between how good we think we are at spotting fakes and how good we actually are.
5. Only 0.1% of people could reliably identify every deepfake shown to them.
One in a thousand. In iProov's 2,000-person study, essentially everyone missed at least one fake, and most missed many. Whatever your mental image of a deepfake looks like, the good ones don't look like it.
6. Over 60% of consumers are confident they can spot a deepfake; actual accuracy on quality video is 24.5%.
Worse than a coin flip, delivered with confidence. This overconfidence gap is the psychological engine of deepfake fraud: people who believe they can't be fooled don't verify.
7. In standardized testing, average detection scores were 0.07–0.08 on a scale where 0 is random guessing.
UK, US, and Brazilian participants all scored barely above chance. The polite research phrasing is that humans are not a reliable detection layer. The practical phrasing is that your eyes are no longer evidence.
8. Voice cloning needs as little as 3 seconds of audio to produce an 85% voice match.
Three seconds is one sentence of a voicemail greeting, one Instagram story, one "hello?" on a spam call. This is why voice has become the fastest-growing deepfake channel, and why callback verification on a known number beats trusting your ears.
The deepfake fraud numbers
Fraud is where the growth curve turns into money. The through-line: deepfakes went from exotic to routine in about three years.
9. A deepfake fraud attempt occurred about every 5 minutes in 2024.
Not every five days. Every five minutes, one identity-verification provider's network alone caught a deepfake attempt. Whatever share of attacks that network sees, the true global rate is higher.
10. Deepfake fraud attempts surged ~3,000% in 2023, including +1,740% in North America.
The breakout year. Over three years, deepfakes went from about 0.1% of detected fraud to roughly 6.5%, about 1 in every 15 fraud cases, and they now feature in around 40% of biometric fraud attempts.
11. 62% of organizations experienced a deepfake incident in the past 12 months.
Regula's global survey tells the same story from another angle: roughly half of companies have been targeted with audio (41%) or video (35%) deepfakes. This is no longer a threat briefing slide; it's a majority experience.
12. The average deepfake identity-fraud attempt costs businesses about $450,000–500,000; large firms average $680,000.
Per attempt. And the ceiling is far higher: a finance employee at engineering firm Arup wired $25 million after a video call where every colleague on screen, including the CFO, was synthetic. He was the only real person in the meeting.
13. The FBI logged $893 million in reported US losses from AI-assisted scams in 2025.
Across 22,364 complaints with a reported AI nexus: $632M from investment scams, $30M from business email compromise, and $19M from AI-assisted romance scams, the deepfake upgrade to a fraud category that already costs Americans over a billion dollars a year, covered in my online dating statistics.
The human victims
Fraud gets the headlines, but by volume, deepfakes are overwhelmingly used for something else, and it's worth saying plainly.
14. 96–98% of deepfake videos online are non-consensual intimate content.
Not politics, not comedy, not fraud. The dominant use of this technology, by an enormous margin, is fabricating explicit imagery of people who never consented to it.
15. Roughly 99% of the victims of sexual deepfakes are women.
Celebrities and private citizens alike: classmates, coworkers, exes. A growing number of US states and countries now criminalize non-consensual deepfake imagery, and platforms are required to remove it in more jurisdictions each year. If it happens to you or someone you know, document it and report it; takedown routes exist and are improving.
16. Celebrity deepfake incidents jumped 81% in a single quarter; politicians were impersonated 56 times in Q1 2025 alone.
Public figures are the free training data of the deepfake economy: endless footage, endless audio. The same mechanics power the synthetic-influencer fraud problem, $2.1 billion in fake-profile losses, which I break down in my AI influencer statistics.
The preparedness gap and the arms race
Here's the strangest pattern in the whole dataset: everyone is getting hit, and almost nobody has a plan.
17. Over 80% of companies have no formal protocols for deepfake attacks, and only 10% prioritize deepfake recognition in training.
Meanwhile 1 in 4 business leaders admit they don't really understand what deepfakes are. The gap between 62% exposure and 10% training is where the $450,000 average losses live.
18. Gartner projects 30% of enterprises will no longer trust identity verification on its own by 2026, because of deepfakes.
Think about what that means: face-matching, the technology banks and governments spent a decade deploying, is being aged out of sole-source trust by synthetic faces. Verification is going multi-layered, biometrics plus liveness plus behavior, because no single check survives.
19. The deepfake detection market is racing toward $15.7 billion by 2026, up from $5.5 billion in 2023.
Growing about 42% a year. Detection is now one of the fastest-growing categories in security, and platforms are watermarking at the source: Google's SynthID alone has marked over 10 billion pieces of content.
20. The deepfake generation market is under $1 billion, an 18× spending asymmetry.
The defender's dilemma in one ratio: the world spends roughly eighteen times more detecting synthetic media than the market for creating it is worth. Deloitte projects US losses from generative-AI-enabled fraud could reach $40 billion by 2027. Defense is expensive; attack is nearly free.
This data is why I tell readers: never hand your real face and voice to an app you don't trust.
Three seconds of audio clones a voice. A few photos clone a face. That is exactly why my standing rule for AI companion platforms is to never use your real name, photos, or identifying details, and why I test how platforms handle data, with my own accounts and my own money. The market numbers are in my AI girlfriend statistics, and the apps are compared on my homepage.
Read the AI girlfriend statistics →Conclusion
The deepfake statistics of 2026 describe a technology in full sprint: half a million files to eight million in two years, a fraud attempt every five minutes, a $25 million video call, and a detection industry burning toward $15.7 billion while over 80% of companies still have no plan.
My read as a data researcher: the single most important number here is 24.5%. That's how often people correctly identify a high-quality deepfake video, and it means "seeing is believing" quietly stopped being true sometime in the last few years. The world hasn't absorbed that yet, and the gap between the 60% who think they'd spot the fake and the 0.1% who actually can is exactly where the next few billion dollars of losses will come from.
FAQs
How many deepfakes are made each year?
Roughly 8 million deepfake files were projected to circulate online in 2025, up from about 500,000 in 2023, implying millions of new deepfakes created per year at a growth rate near 900% annually. Precise counts are impossible; these are the estimates cited by Europol and the UK government.
What is a deepfake in AI?
A deepfake is synthetic media, video, audio, or images, created or altered with AI to show a real person saying or doing something that never happened. The name combines "deep learning" and "fake". The same underlying generative technology also powers legitimate products, from film effects to virtual influencers and AI companion apps.
How do deepfakes work?
At a high level, AI models are trained on images, video, or audio of a person until they can generate new, convincing footage or speech of that person. Modern tools need surprisingly little input: a few photos for a face, or as little as 3 seconds of audio for an 85% voice match. That low barrier to entry is why volume grew 16-fold in two years.
Can you detect a deepfake?
Not reliably by eye: humans are 24.5% accurate on high-quality deepfake video, and only 0.1% of people catch every fake. Practical defenses are procedural instead: verify unusual requests through a second channel, call back on a known number, use code words with family, and treat urgency plus a money request as a red flag no matter how real the voice or face seems.
What percentage of deepfakes are pornographic?
An estimated 96 to 98% of deepfake videos online are non-consensual intimate content, and about 99% of the victims are women. This remains the dominant use of the technology by volume, and a growing number of jurisdictions now criminalize creating or sharing it.
Sources
- iProov: Deepfake detection study and statistics. iproov.com
- Eftsure: Deepfake statistics for CFOs (DeepMedia, Regula, Gartner, Entrust data). eftsure.com
- Keepnet: Deepfake statistics and trends (Sumsub, Pindrop, Arup case). keepnetlabs.com
- StationX: Deepfake statistics (Sensity, Onfido, market data). stationx.net
- DeepStrike: Deepfake statistics: the AI fraud wave. deepstrike.io
- SQ Magazine: Deepfake statistics (growth, incidents, regions). sqmagazine.co.uk
- StingRai: Deepfake statistics 2026 (Europol IOCTA, Gartner, Veriff/Kantar). stingrai.io
- Memeburn: FBI IC3 2025 AI-nexus loss data. memeburn.com
- Bright Defense: Deepfake statistics (Deloitte, Pindrop, Google SynthID). brightdefense.com
- UNESCO: Deepfakes and the crisis of knowing. unesco.org