Deepfakes Explained: How They Work, How To Spot Them And Why They Are Becoming Dangerous

Can You Still Trust What You See? The Deepfake Revolution Explained

How AI Can Steal A Face, Voice And Identity

For most of modern history, photographs, recordings and video carried an assumption that written claims did not: somewhere, something had actually happened in front of a camera or microphone. Artificial intelligence is rapidly weakening that assumption. A person can now appear to say words they never spoke, make facial expressions they never made or seemingly appear in footage that never existed.

Deepfakes are therefore becoming more than an internet novelty. They are colliding with fraud, identity theft, sexual abuse, political manipulation and cybersecurity at the same moment that generative AI is making synthetic media faster and easier to produce. The central problem is no longer simply whether AI can create something fake. It is whether ordinary people can reliably work out what is real.

What Exactly Is A Deepfake?

A deepfake is artificially generated or manipulated media designed to resemble something authentic. It can involve someone's face being placed onto another person's body, their expressions being altered, their lips synchronised to fabricated speech, their voice cloned or an entirely synthetic version of them being generated.

The name combines "deep learning" with "fake", although modern deepfakes do not rely on one single technique. Research now covers face swapping, facial reenactment, talking-face generation and facial attribute manipulation, while newer diffusion-based systems have joined earlier approaches based on autoencoders and generative adversarial networks.

That distinction matters because the stereotypical deepfake — somebody crudely attaching one celebrity's face to another person's video — describes only part of the modern threat. Synthetic media can now involve images, moving video and audio, and the pieces can be combined.

A convincing fake video accompanied by an obviously robotic voice will raise suspicion. A convincing fake face combined with convincing cloned speech, realistic lip movements and believable context is far more powerful.

How Deepfakes Actually Work

At the simplest level, a deepfake system learns patterns in real examples and uses those patterns to construct new media. Give a model enough information about a face and it can learn features such as its shape, expressions, skin appearance and how it changes under different poses or lighting conditions.

Earlier face-swapping systems commonly used encoder-decoder architectures, while generative adversarial networks became famous for realistic synthetic imagery. In a GAN, one network generates artificial content while another tries to distinguish generated material from real examples. The competition between the two can drive increasingly convincing output.

Diffusion models have pushed synthetic image generation further. In simplified terms, these systems learn how to reverse a process that gradually turns an image into noise. Once trained, they can begin with noise and progressively construct a coherent image, guided by instructions and other conditioning information.

Modern deepfake techniques can also preserve one person's identity while borrowing movement from somebody else. Facial reenactment can transfer head position and expressions. Lip-synchronisation systems can alter a mouth to correspond with new speech. Face-swapping systems can replace one identity with another while attempting to preserve the original scene, movement and lighting.

The result is less like traditional video editing and more like reconstruction. Instead of manually changing every frame, AI predicts what the manipulated frame should look like.

Why Modern Deepfakes Look So Convincing

Early deepfakes offered comforting clues. Faces could flicker between frames. Eyes behaved strangely. Teeth appeared distorted. Lighting did not match. Skin could look unnaturally smooth and the boundary between the manipulated face and the surrounding head was sometimes obvious.

Those clues still appear, especially in poor-quality fakes, but relying on them is increasingly dangerous.

Modern generation systems are better at producing realistic textures, coherent facial structure and convincing lighting. They can also produce completely synthetic imagery rather than merely pasting one face onto another. NIST now describes realistic deepfake creation as a widely accessible, low-cost process and is testing detectors against face swapping, body swapping, contextual manipulation and deliberate adversarial attacks designed to make detection harder.

Social media creates another advantage for the faker. A video does not necessarily need to survive forensic inspection at full resolution. It may only need to look convincing for several seconds on a phone.

Compression, reposting, cropping and low resolution can hide abnormalities that would otherwise expose manipulation. The conditions under which millions of people consume online media can therefore help conceal the imperfections in synthetic content.

Voice Cloning Changes The Threat

Deepfakes do not have to involve video at all.

AI voice cloning can model the characteristics of a person's speech and create new audio that sounds as though the target said something they never said. The technology has legitimate uses, including helping people who have lost the ability to speak, but the same capability can be used for impersonation.

The danger comes from the psychological power of a familiar voice. A strange email asking for money may trigger suspicion. Hearing what appears to be your son, daughter, partner, manager or company director asking for urgent help can bypass that scepticism.

Consumer protection authorities have warned that voice-cloning programs can work from relatively short samples of recorded speech. Those samples may already exist publicly in social-media videos, interviews, podcasts, presentations or other online material.

This changes the old rule that hearing somebody provides evidence of who is speaking. Increasingly, it does not.

How To Spot A Deepfake

There is no single visual trick that reliably identifies every deepfake. Anyone claiming otherwise is offering a level of certainty the technology no longer supports.

Poor synthetic video can still reveal itself through inconsistent lighting, unusual reflections, distorted hands or teeth, unnatural mouth movements, strange facial boundaries, unrealistic blinking, temporary changes in facial detail or inconsistencies between frames. Audio can contain odd pacing, unnatural intonation, abrupt changes in background noise or speech that does not sound quite like the person's normal vocabulary and manner.

Context is often more useful than pixels.

Ask whether the footage has a credible origin. Look for the original recording rather than relying on a clipped repost. Be suspicious when shocking material appears through an anonymous account with no verifiable source. Check whether the person's supposed actions make sense and whether independent evidence exists.

For financial requests, the most important question is not "Does this sound real?" It is "Can I independently verify who is asking?"

If someone who appears to be a relative, executive or colleague suddenly requests money or sensitive information, contact them through a number or channel you already know. Do not rely on contact details supplied inside the suspicious message itself. That independent verification is increasingly more dependable than trying to outsmart a sophisticated synthetic voice.

Why Automated Detection Is Getting Harder

AI can be used to detect AI, but this creates an arms race.

Detection systems may search for visual inconsistencies, frequency patterns, generation artefacts, abnormal audio signals or other statistical traces that humans cannot easily perceive. The problem is that a detector trained to recognise yesterday's generation methods may perform badly against tomorrow's.

This gap between laboratory performance and the real world is becoming a major concern. NIST's 2026 deepfake forensic work says current AI detection systems can suffer performance degradation of roughly 45 to 50 per cent when moving from academic evaluation into operational deployment. Its new testing programme deliberately includes adversarial modifications because realistic detection systems must survive attempts to defeat them.

Watermarking and provenance technology may help, but they are not magic solutions either. A generator can embed information showing that content was synthetically produced, while authentication systems can record where genuine media came from.

The difficulty is universality. Not every generator applies the same standards, old content lacks modern provenance information and malicious actors may attempt to remove or damage watermarks. The FTC has warned that watermark-based defences have limitations and can potentially be altered or removed.

How Deepfakes Are Being Used For Fraud

The most immediate deepfake threat may not be a fake presidential speech. It may be somebody pretending to be a person you already trust.

Fraud works by manipulating confidence and urgency. Deepfakes make both easier.

Imagine a finance employee receiving an apparently genuine voice message from a senior executive. A parent gets a call apparently from a distressed child. An employee receives a video message from someone resembling their boss. A victim meets someone online who appears repeatedly on video but whose identity has been artificially constructed.

These scams do not require perfect synthetic media. They only need to be convincing enough for long enough to push the victim into acting.

Authorities are already seeing that pattern. The FBI has warned that malicious actors have used AI-generated voice messages while impersonating senior US officials, with targets encouraged to move communications onto other platforms before attempts to obtain information, access accounts or elicit funds.

The key weakness being exploited is human trust. Cybersecurity has spent decades trying to secure machines. Deepfakes increasingly target the person operating the machine instead.

The Political And Disinformation Threat

Political deepfakes attract attention for an obvious reason: a fabricated recording released at the right moment could make a candidate appear to confess, insult a group, announce a policy or react to an international crisis.

But the political danger extends beyond persuading millions of voters that one specific fake is genuine.

Synthetic media can overwhelm an information system with uncertainty. Thousands of misleading images, recordings and short videos can be created more cheaply than professionally investigating them. Verification takes time; fabrication can be almost instantaneous.

That imbalance benefits anyone who wants to create confusion.

It also creates the reverse problem. Once the public knows convincing deepfakes exist, genuine evidence can be dismissed as artificial. A real recording of misconduct can be answered with a simple claim: "It's AI."

This is sometimes described as the liar's dividend. The existence of fake evidence can weaken confidence in real evidence.

The long-term political danger is therefore not merely believing falsehoods. It is reaching a point where people stop believing anything.

Non-Consensual Deepfakes Have Become A Legal Flashpoint

Some of the most personally destructive deepfakes involve fabricated intimate imagery.

AI systems can take ordinary photographs of a real individual and generate sexually explicit images that appear to depict them. The victim does not need to have taken an intimate photograph in the first place. Their identity itself becomes the raw material.

UK law has moved significantly in response. New offences brought into force in February 2026 criminalised creating, or requesting the creation of, purported intimate images of adults without consent or a reasonable belief in consent. Further provisions that came into force in June 2026 target the making or supply of tools intended to generate purported intimate imagery.

The legal shift reflects something important about the technology. A deepfake can be false as evidence while producing completely real consequences for its target.

Reputational damage, humiliation, harassment and loss of privacy do not disappear simply because the image was generated by a computer.

Governments Are Starting To Demand Transparency

Regulators are increasingly moving from debating synthetic media to imposing specific obligations.

The European Union's AI Act provides one major example. Article 50 transparency requirements became applicable on 2 August 2026. They include requirements concerning machine-readable marking of AI-generated or manipulated material and disclosure obligations for deepfake content in relevant circumstances.

The goal is straightforward: people should have a better chance of knowing when apparently authentic material has been artificially generated.

Yet regulation faces the same structural problem as detection. Synthetic media is global. Someone producing malicious material may operate outside the jurisdiction in which the victim lives, distribute it anonymously and move it through multiple platforms within minutes.

Law can increase consequences and force responsible companies to adopt safeguards. It cannot restore a world in which manipulated media is technically difficult to make.

How To Protect Yourself

The most important defence against deepfakes is to change what counts as proof.

A familiar voice is no longer enough. A recognisable face is no longer enough. A video apparently showing someone speaking is no longer enough. High-stakes requests should increasingly require a second channel of verification.

Families can agree on a private word or verification method for genuine emergencies. Businesses can require additional approval for unusual transfers regardless of who apparently requested them. Individuals should contact people directly through previously verified numbers when money, passwords or confidential information are involved.

Urgency should itself raise suspicion. Fraudsters want victims acting emotionally before verification catches up.

The same principle applies when consuming sensational online content. If a video would dramatically alter your opinion of somebody, that is exactly when verification matters most.

The Fight Over Reality Has Only Just Begun

Deepfakes are not dangerous because every fake is perfect. They are dangerous because they are becoming cheap, scalable and convincing enough to exploit the brief moments in which human beings decide what to believe.

Detection technology will improve. Provenance systems will spread. Platforms will develop stronger safeguards and governments will write new laws. Generation systems will improve at the same time.

That leaves society confronting a deeper change. For generations, cameras and microphones strengthened the credibility of evidence because manufacturing convincing recordings was difficult. Generative AI is destroying that scarcity.

The new challenge is not learning to distrust everything we see. A society that reaches that point has already lost something valuable. The challenge is building new methods of authentication strong enough that truth can remain provable even when convincing fiction becomes almost effortless to manufacture.