Deepfakes and Fake Media

When seeing is no longer believing

For most of history, a photograph or a video clip counted as strong evidence that something really happened. Cameras could be staged and pictures could be edited, but convincingly faking a moving, talking human was slow, expensive and rare. That assumption is now breaking down. Deepfakes are synthetic media, usually video or audio, in which artificial intelligence makes a real person appear to say or do something they never said or did. The name combines "deep learning", the kind of AI behind them, with "fake".

This lesson is not about panic. It is about understanding a genuinely new situation clearly enough to navigate it. The same family of tools that lets you generate playful pictures, explored in Generative AI: Images and Text, can also be turned to deception. Knowing how that works, what it can and cannot do, and how to respond is now a basic part of media literacy.

How a deepfake is actually made

Deepfakes are built on the same core idea as other generative AI: a model is trained on many examples until it learns the patterns well enough to produce convincing new examples of its own. For a face-swap video, a system studies many images of a target person from different angles and expressions, learning how their face moves and lights. It can then map that learned face onto someone else's performance, frame by frame, so the target appears to be the one talking.

A particularly influential approach has been the generative adversarial network, or GAN. Here two neural networks are trained in competition. One, the generator, tries to produce fake images; the other, the discriminator, tries to tell fakes from real ones. Each round, the generator gets better at fooling the discriminator and the discriminator gets better at catching it. The result of this contest is a generator that produces strikingly realistic output. Other methods, including diffusion models, now do similar work. The deeper mechanics of these networks are covered in Neural Networks Explained.

Voice is just as vulnerable. Voice cloning tools can learn the sound of a particular person from a surprisingly short sample, sometimes only a few seconds, and then read any text in that voice, complete with their accent and rhythm. Combine a cloned voice with a synthetic face and you have a fake video of a named person saying scripted words.

Why this matters: the real harms

It is tempting to file deepfakes under "cool but harmless tech". The harms are real and already documented.

Fraud and scams. Criminals have used cloned voices to impersonate a relative in distress or a company boss, phoning to demand an urgent money transfer. Hearing a familiar voice lowers people's guard, which is exactly the point.

Harassment and abuse. A large share of deepfakes online are non-consensual fake intimate images, overwhelmingly targeting women and including minors. This is a serious form of abuse, and many places have passed or are passing laws against it. Putting someone's face into content they never consented to is harmful regardless of how "realistic" it looks.

Political manipulation. A fabricated clip of a leader announcing something false, released at the right moment, can spread faster than any correction. Even if it is debunked within hours, the false impression can linger and the timing can do damage.

The liar's dividend. This is the subtler harm, and arguably the most corrosive. Once everyone knows convincing fakes exist, a person caught on genuine footage can simply claim it is a deepfake. The mere possibility of fakery gives the dishonest a way to dodge real evidence. The danger is not only that we believe false things, but that we stop trusting true things. A society where no recording can be trusted is a society easier to lie to.

Spotting fakes: useful but not enough

People often ask for a checklist of visual giveaways: unnatural blinking, blurry edges where a face meets hair, lighting that does not match, lips slightly out of sync, strange hands or jewellery. These tips have some value, and on weaker fakes they work. But there are two honest limits.

First, the technology improves quickly. Many of yesterday's reliable tells, like the famous "deepfakes don't blink properly" trick, were fixed soon after they became known. Betting on visible glitches is betting against a moving target.

Second, automated detection tools exist, but they are caught in an arms race. Every detector becomes training material for making the next generator harder to detect. Detection helps and is worth doing, especially at scale, but it will always lag behind and can be evaded. Treat it as one layer, not a guarantee.

One promising defensive direction is provenance: cryptographically signing media at the moment of capture so its origin and any edits can be verified later. Industry standards for this are emerging. They do not prove something is true, but they help establish where a piece of media came from, which is often the real question.

A better habit: check the source, not just the pixels

Because you cannot reliably eyeball every fake, shift your effort from detecting fakery to verifying provenance. Practical questions that work even when a fake is technically flawless:

1. Where did this first appear? Trace it back. A clip with no clear original source deserves suspicion.
2. Do trusted, independent outlets report the same thing? A genuinely major event is rarely covered by only one anonymous account.
3. Does this conveniently confirm what someone wants me to believe? Fakes are designed to ride existing anger and tribal loyalty. Content that makes you feel instantly furious is exactly what is most worth faking, and most worth pausing on.
4. What is the motive and timing? A shocking clip dropped right before an election or during a crisis warrants extra scepticism.
5. Slow down before sharing. Speed is the friend of misinformation. A few minutes of checking breaks the chain.

Living in a world with synthetic media

It would be wrong to end on doom. Deepfake technology is not purely a weapon. The same techniques dub films into dozens of languages while matching lip movements, recreate the voice of someone who has lost theirs to illness, bring historical figures to life in museums, and power clearly labelled satire and art. The technology is neutral; consent, honesty and context decide whether a given use is fine or harmful.

What changes is the default assumption. We are moving from "video is proof" to "video is a claim that needs checking", much as we already treat text and photos. That is not the end of truth. It is a demand for better habits: valuing trustworthy sources, supporting provenance standards, and refusing to be the link that spreads a lie. The deeper questions of who is responsible for these systems and how to keep them fair run through AI Ethics and Fairness, and they apply squarely here. Your scepticism, used wisely, is part of the defence.