Deepfakes and synthetic media

For most of history, a photo or a video was decent evidence that something happened. That era is over. AI can now generate faces, voices and whole scenes that never existed, and it can put real people into situations they were never in. This lesson is about understanding that shift and knowing what to do about it.

What a deepfake is

A deepfake is synthetic media: an image, video or audio clip generated or altered by AI to show a real person doing or saying something they never did. Early deepfakes were glitchy and easy to laugh off. Today's are not. The tools have become cheap, fast and widely available, and the results can fool careful viewers, including experts. It no longer takes a studio or special skills. A single public photo or a few seconds of video can be enough source material.

Why the fakes got so good

So how does a machine put a real face into a video that never happened? The short answer is pattern learning. A model is shown enormous numbers of faces until it learns the underlying rules of what faces do: how skin catches light, how a mouth moves through a word, how a head turns and settles. Once it knows the general rules, it needs surprisingly little of any one person, because your face is just a particular setting of patterns it already understands. The model is not copying your photos. It is describing you in a language of faces it already speaks fluently, which is why a handful of images is enough.

The same idea explains why the fakes improved so quickly. A decade ago this work needed research labs and serious computing power. Since then the models have become better, the computing has become cheaper, and the whole pipeline has been packaged into apps that need no skill at all. Anything that becomes cheap and easy becomes common, so the sensible assumption now is that any image, video or voice you meet online could be synthetic, and the real question is how to work out which ones matter.

The harm is real

Some synthetic media is harmless fun, clearly labelled and made with consent. The serious harm comes when it is neither. Deepfakes have been used to spread fake political statements, to fake celebrity endorsements for scams, and, most damaging of all, to create sexualised images of real people without their consent. The targets are often ordinary people, including school students, and the damage to them is real even though the image is fake.

Be very clear on the law here. In Australia, creating or sharing sexualised deepfakes of a real person is a crime, and that includes doing it as a joke, and it includes images of classmates. It is treated as image based abuse, and both the criminal law and the eSafety Commissioner take it seriously. If it happens to you or to someone you know, it is not your fault and there is a clear path to get help.

There is also a quieter harm that runs in the opposite direction. When everyone knows video can be faked, a person caught doing something real can simply claim the footage is a deepfake, and some people will believe them. Researchers call this the liar's dividend: the mere existence of convincing fakes gives cover to liars even when nothing has been faked at all. So deepfakes damage trust in both directions, making false things believable and true things deniable, and the second harm may prove the larger one over time.

A misconception worth clearing up

Some students believe that only the person who makes a harmful deepfake is in legal trouble, and that forwarding one is harmless because you did not create it. That is wrong. Sharing image based abuse is itself an offence, and every forward widens the circle of harm for the person in the picture. Federal law was strengthened in 2024 to deal with sexually explicit deepfakes specifically, so this is not a grey area waiting for the rules to catch up. The rules exist now, and they apply to sharing just as much as creating.

Spotting the signs, and why spotting is not enough

There are visual tells worth knowing: hands and fingers that look wrong, jewellery or glasses that warp between frames, lighting that does not match the scene, blinking that looks off, audio slightly out of step with lips, and backgrounds that smear when someone moves. Checking for these is a good habit. But be honest about the limit: the tells shrink with every new model generation, and a careful fake may show none of them. You cannot rely on your eyes alone.

You might hope a detection tool could do the checking for you, and detectors do exist, however they sit inside an arms race they are built to lose. A detector learns the fingerprints of today's fakes, and the moment those fingerprints are known, the next generation of generators can be trained to avoid leaving them. Detection also fails in both directions: it misses careful fakes and it flags genuine footage as suspicious, and a tool that is wrong in either direction cannot settle an argument. Treat any detector's verdict as one weak clue, never a ruling.

That is why the deeper defence is about source, not pixels. Ask where the media came from, who posted it first, and whether anyone reputable has confirmed it. The verification lesson later in this topic builds that habit properly.

What this looks like in your week

The clip of a politician saying something outrageous that lands in the group chat, the celebrity giveaway video a relative shares on Facebook, the edited image of a classmate that someone insists is just a joke. Each of these calls for the same two moves. First, do not forward it, because forwarding is how synthetic media does its work, and you become part of the machinery the moment you press share. Second, ask the source question before the content question: who posted this first, and how would they know. And if the media targets a real person you know, treat it as the serious matter it is rather than content to react to, because for the person in the image the fakeness changes nothing about how it feels.

This lesson and the next two form one build. Here you have seen that media itself can be fabricated. The next lesson shows the same tools aimed at your family's money, and the lesson after that covers what AI services quietly collect from you. The final lesson then teaches the verification habits that answer all three, because the right response to a synthetic internet is a small set of checking habits used consistently, not fear.