Deepfake Statistics: What the Numbers Actually Show
Numbers in this field are quoted confidently and sourced badly. Before the figures, a warning about them — then what the evidence does support, and our own data, which is small but at least ours.
Why most quoted numbers are unreliable
Three problems recur in almost every statistic you will see cited.
Detection bias. Counts of deepfakes are counts of *detected* deepfakes. A better fake is an uncounted one, so every figure is a floor rather than a measure.
Vendor incentive. Many widely quoted numbers come from companies selling detection. That does not make them false, but it does mean the methodology deserves reading before the headline gets repeated.
Ageing fast. A 2023 figure describes a different technical world. Cost, quality and volume all changed substantially since.
So treat any specific percentage, including anything below, as an indication of direction rather than a measurement.
What the evidence does support
The overwhelming majority of deepfake material is non-consensual intimate imagery. Every serious survey of the field has found this since the first ones in 2019, and the proportion has stayed high. Political deepfakes dominate coverage and are a small minority of actual volume.
Targets are mostly private individuals. Public figures make the news; ordinary people make up the bulk of victims, and have far less recourse.
Fraud is the fastest-growing category. Voice and video cloning used in business payment fraud, fake job interviews and romance scams. This is where the money is, and where the growth is.
The cost collapsed. What needed a large photo set and days of training now runs from a handful of images, sometimes live. That is the single most important fact in this whole area.
Our own numbers
Small sample, but ours and honestly reported: of every check run through FakeSec since launch, roughly four in ten come back as AI-generated, half as authentic, and about one in eleven as inconclusive.
That last figure is the one we would highlight. Inconclusive is not a failure — it is what an honest tool returns when its methods genuinely disagree, and any detector reporting near-zero uncertainty is hiding something. The live breakdown sits on the main page, updated automatically.
The 40% figure also reflects who bothers to check: people who already suspect something. It is not a measure of how much of the internet is generated.
What to do with any of this
The useful conclusions from the data are behavioural, not statistical.
Video calls are no longer identity proof — ask for occlusion and profile turns. Intimate images sent to anyone can be weaponised, and fabricated ones require no cooperation from you at all. Payment requests confirmed “by video” need a second channel of verification.
If you are dealing with a specific file, our deepfake detector handles photos and frames, and the spotting guide covers what to check by eye.
Key points
- Deepfake counts measure detection, not prevalence — every figure is a floor.
- Most deepfake material is non-consensual intimate imagery, not political.
- Private individuals are the majority of targets.
- Fraud is the fastest-growing use, driven by collapsed production cost.
- Of our own checks: ~40% AI, ~50% authentic, ~9% inconclusive.
Frequently asked questions
Check an image yourself
Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.
More from the blog
Try it on your own image
Free account, 3 checks included, no card required.