How deepfakes are made, why humans are fooled, what voice cloning can do and what AI-generated media means for trust, consent and evidence. Understanding the mechanism is the first step to not being deceived by it.
The AI needs hundreds or thousands of images or video frames of the target person. Every angle, lighting condition and expression helps. Public figures are most at risk because their images are widely available online.
A neural network converts each face into a compact numerical description — a vector representing the unique geometry of that face. Eyes, nose, jaw, skin texture, proportions. All reduced to numbers.
A generative model learns to reconstruct the target face in any pose, lighting or expression — by comparing its output against real images and adjusting until it gets it right. This is the GAN (Generative Adversarial Network) process.
The trained model replaces one face with another in video frames, or generates entirely new face images that never existed. The result is processed to match lighting, skin tone and motion blur of the original.
Edges are blended. Lighting is corrected. Temporal consistency (frame to frame smoothness) is enforced. The result can be indistinguishable to human eyes in casual viewing.
Human brains have dedicated regions for face recognition. We are wired to read emotion, intent and identity from faces quickly and automatically. This system was not built to detect statistical forgeries.
For most of human history, a photograph or video was evidence that something happened. That assumption is now broken — but our instinct to trust visual evidence remains very strong.
In normal viewing we do not analyse every pixel. We get an impression. Deepfakes exploit this — they only need to fool a fast impression, not a forensic inspection.
When video is compressed for social media, many deepfake detection clues — unnatural blinking, edge blurring, texture inconsistencies — are destroyed. The platform itself covers the forgery's tracks.
Modern voice cloning models can clone a voice from a very short sample. A phone call, a voicemail, a social media video. The model extracts vocal characteristics — pitch, cadence, accent, timbre.
Once cloned, the attacker can type any text and the model will speak it in the target's voice. The target never said those words. The voice is entirely synthetic.
Legitimate: restoring speech for people who have lost their voice. Misuse: fraud phone calls, fake ransom calls, fake celebrity endorsements, impersonating family members to steal money.
AI-generated fake intimate images of real people — most often women — produced without consent. Now illegal in the UK under the Online Safety Act 2023.
Cloned voices used in phone scams. Victims receive calls that sound exactly like a family member claiming to be in trouble and needing money immediately.
Cases have been reported of deepfake images of classmates being created and shared within schools. This is a form of sexual harassment and is illegal.
Fake videos of politicians, journalists and public figures saying things they never said — used to spread false information or damage reputations.
Synthetic media used as fake evidence in disputes, investigations or court cases. The legal system is still developing ways to verify digital evidence.
Fake images or videos used to gaslight, blackmail or emotionally control victims. "I have footage of you" — when no such footage exists.
Early deepfakes blinked rarely or unnaturally. Newer models are better but eye movement can still look slightly wrong.
Light hitting the face from a different angle than the background. Shadows that do not match the scene.
AI still struggles with hands (wrong number of fingers, unnatural positions) and teeth (blurred or merged).
The boundary between the face and hair or background can appear slightly blurred or unnaturally smooth.
Earrings, glasses and facial hair are difficult for AI to render consistently. Look for morphing or flickering.
Lip movement that does not perfectly match the audio. Slight delays or mismatches — particularly noticeable on certain sounds.
Search the image or a frame from the video. If it appears elsewhere in a different context, it may have been manipulated.
Does this content serve a purpose? Who benefits from you believing it? Emotional or shocking content is more likely to be faked to provoke a reaction.
SCENARIO: Strange phone call You receive a call. The voice sounds exactly like your parent or sibling. They say they are in trouble, in a foreign country, and need you to send money immediately. They beg you not to tell anyone else yet. // What AI is doing Voice cloned from: social media videos, voicemails, calls Time needed to clone: as little as 3 seconds of audio Script: typed by the attacker, spoken by the clone Goal: urgency + secrecy = bypassing your critical thinking // What to do 1. Hang up 2. Call that person back on their known number 3. Ask them a question only they would know the answer to 4. Never send money or share details based on a phone call alone // The rule If a caller creates urgency and asks for secrecy — that is the scam. Real emergencies can wait 60 seconds for you to verify.
SCENARIO: Shocking video A video appears on social media showing a well-known person saying something outrageous, offensive or criminal. It is spreading fast. People are reacting with anger. // What AI is doing Face mapped onto a different body or existing video Voice cloned or replaced with text-to-speech Platform compression: destroys detection artefacts Emotional content: designed to spread before anyone checks // What to do 1. Do not share before verifying 2. Slow down — emotional reaction is part of the design 3. Check the source — where did this first appear? 4. Reverse image search the thumbnail 5. Look for official responses or fact-checker coverage 6. Check for blurring around the face and hair edges 7. Watch for unnatural blinking or lip sync issues // The rule The more outrageous and emotional the content, the more carefully you should check it before believing or sharing it.
SCENARIO: Image of someone you know You are shown an image claiming to be of someone you know — a friend, classmate or family member — in a compromising situation. The image looks realistic. // What AI is doing Face swapped onto a different body using deepfake technology Or: entirely generated from reference photos of that person Cost to produce: potentially very low with available tools Intent: harassment, blackmail, humiliation, control // What to do 1. Do not share the image under any circumstances 2. Tell a trusted adult or contact police 3. In the UK: non-consensual intimate deepfakes are illegal (Online Safety Act 2023 — sharing carries criminal penalties) 4. The person in the image is a victim — treat them accordingly 5. Screenshot evidence of who sent it and when // The rule Creating or sharing fake intimate images of real people is illegal. The victim is never at fault. The person who created or shared it is.
SCENARIO: News clip A news clip appears showing a politician, celebrity or official making a statement that seems to contradict everything they have previously said or done. // What AI is doing Either: genuine clip taken out of context Or: deepfake video replacing face/voice Or: manipulated audio (cheaper than full video deepfake) Or: real clip from different country presented as local // What to do 1. Find the original source — not just a repost or screenshot 2. Check multiple news outlets — if only one source has it, be sceptical 3. Look for original timestamp and context 4. Check the official account of the person involved 5. Use a fact-checking service // The rule Authentic news can be verified through multiple independent sources. Fake news relies on you not checking.
The assumption that photographs and videos represent reality is now broken. This is a fundamental shift in how we need to evaluate information. Critical checking is now a necessary skill, not optional.
The UK Online Safety Act 2023 criminalised sharing non-consensual intimate deepfakes. More legislation is in progress. The technology moved faster than the law but the law is following.
Using someone's likeness or voice without permission — especially to create intimate content — is a violation regardless of the technology used to do it. The digital nature does not reduce the harm.
Voice restoration for people who have lost speech. Historical reconstruction. Accessibility tools. Film production. Translation dubbing. The technology is not evil — misuse without consent is.