Deepfakes: what they are and how they are detected
Faces swapped. Voices cloned from seconds of audio. Video of people who never said what they are saying. What deepfakes are used for in fraud, and how they are detected.
Updated 2026-09-11 · Veriflied
What the word covers
Deepfake broadly covers synthetic depictions of real people: face swaps in images and video, driving a real face with someone else's expressions (reenactment), fully synthetic persons, and voice cloning from short samples. What they share is imitation of a specific, real person — that is what separates deepfakes from ordinary AI images.
Where it hits
CEO fraud
Cloned voices and video calls used to make employees transfer money. Cases exist across Europe, and the amounts are large.
Identity and KYC
Face swaps against ID photos and liveness checks in account opening and loan applications.
Romance scams
Profile pictures and video calls with borrowed or synthetic faces.
Extortion and misinformation
Fabricated recordings of real people, private and public.
What you can do yourself
On video calls: ask the person to turn their head fully sideways, pass a hand in front of their face or remove their glasses — profiles, occlusions and fast movement remain weak points of the technique. On audio: hang up and call back on a number you know. On images: look at the transition between face and hair/neck, ears and reflections in glasses, and run a reverse image search on the profile photo.
Most importantly: agree verification procedures BEFORE it happens. A family code word and a fixed call-back rule at work beat any detector, because they do not depend on being able to see anything.
How technical detection works
Image and video detection looks for blending boundaries around the inserted face, inconsistent lighting and skin texture between face and surroundings, and in video: unnatural expressions and missing micro-movements. Voice detection analyses frequency patterns cloned voices still deviate on. Like all AI detection it is a race: models are retrained as new generators appear, and serious detection always combines several signals rather than one.
Frequently asked questions
Can a deepfake video be exposed with certainty?
Often, but not always. Quality varies enormously, and detection yields a probability with reasons, not a guarantee. Procedures (call-backs, code words) therefore remain the strongest protection in practice.
How little audio does voice cloning need?
A few seconds are enough for a recognisable clone with modern tools. Assume any publicly available voice can be cloned.
Does VeriFly catch deepfakes?
VeriFly is trained against AI-generated and AI-edited images and catches diffusion-based faces and manipulations in image material. Dedicated video and voice detection is a separate track we develop as demand in insurance and KYC grows.
Try it on your own document
Create an account and analyse your first 10 images or documents for free. You get a report with the concrete findings, not just a score. All processing in the EU.