Deepfakes: what's real, what isn't, how to tell

Before we say anything, let's look at some. Pick one of the three demos below and project it at the front. Give the class ten seconds per image to shout 'real' or 'fake' before you reveal the answer.

All three demos are free, browser-only, and need no login.

In March 2023, an image spread across social media of Pope Francis strolling through Rome in a floor-length white Balenciaga puffer jacket. It looked stylish, unusual, and entirely real. Millions of people believed it. The Pope, of course, was wearing no such thing. The image had been generated by a free AI tool called Midjourney by an American construction worker experimenting on a Sunday afternoon.

That was three years ago. Deepfake tools have moved fast since. A short video clip of anyone saying anything can now be generated by any teenager with a laptop and roughly ten minutes. A voice clone of a family member is cheap and quick. And the pixel-level tells (the strange hands, the melted ears, the weird blinking) are being patched with every new model release.

Over the next forty minutes we'll look at how deepfakes actually get made, why the harm scales so fast, the small habits that catch most of what circulates in a school WhatsApp group, and what to do the moment something shocking lands in your feed.

A deepfake isn't a clever edit of an existing image. It's a fresh image (or video, or audio clip) generated from scratch by an AI that was trained on huge piles of real ones.

Here's the shape. Someone feeds an AI hundreds of thousands of photos of faces. The AI learns the statistical patterns: where eyes sit, how skin catches light, how a jaw hinges when someone smiles. Once it's learned those patterns, it can generate a face that has never existed but obeys every rule of what faces look like. The same trick works for voices (trained on hours of speech) and short video clips (trained on frames of people moving).

The two big model families you'll hear named are GANs (used for the early face generators like the ones on thispersondoesnotexist.com) and diffusion models (the newer, sharper approach behind Midjourney, DALL-E and video tools like Sora). The details differ, but the shape is the same: pattern-learning on a huge pile, then generation of something new that fits the pattern.

It helps to know which flavour of deepfake you're dealing with, because the way you spot each one is different.

Before you tap the cards below, have a guess: what are the three main types of deepfake you've already come across, whether you knew it or not?

The three flavours share the same underlying tech, but they land in different places: images circulate on social media, audio arrives by phone, and video is used in fraud and political disinformation.

The stakes with deepfakes aren't only about the fake being believed. There's a second, quieter effect that researchers call the liar's dividend.

Once everyone knows AI can fake anything, anyone caught doing something real on camera can plausibly shrug and say, 'That was a deepfake.' Real footage of real misconduct loses its power. Real evidence in court can be challenged as 'possibly generated'. Real voices of real politicians get dismissed as clones. The problem isn't just that we start believing fake things; it's that we stop believing true ones.

Layer that on top of the scale (billions of images shared daily, most unverified) and the speed (one tool can generate a thousand variations in an hour) and the harm compounds fast. The 2024 US election saw AI-generated content of every major candidate. The Hong Kong CFO scam netted around £20 million in a single meeting. The same technique is showing up in cheaper, closer-to-home ways too: 'family emergency' voice scams targeting parents and grandparents, fake celebrity endorsements pushing dodgy products on TikTok, and deepfaked-teacher clips being shared around school group chats as pranks. And in UK schools, AI-generated intimate images of pupils have already been a safeguarding issue. This isn't a future problem.

You'll read a lot of listicles about 'weird hands' and 'melted ears' and 'unnatural blinking'. Those tells were real in 2023. Most of them have already been patched in the latest models, and the rest will be gone within a year or two.

Before you tap the cards below, have a think: if you had to teach someone in 2028 how to spot a deepfake, what would you tell them to look at?

The first two tells will keep fading with the tech. The last two won't. Build the habits that outlast the model release.

This one is a bit of a trap. It's tempting to paste a suspicious image into ChatGPT or Gemini and ask, 'Is this a deepfake?' The AI will usually give you a confident answer. That answer is often wrong.

Current general-purpose AIs can't reliably detect AI-generated content. There are specialist detection tools that try (Intel's FakeCatcher, various academic projects), but chatbots aren't among them. What chatbots CAN do well is help you run a proper verification workflow: breaking down what an image claims, suggesting searches, flagging visual details worth a second look. The two prompts below show the difference.

Time to put the whole toolkit into practice. It's Sunday evening. A friend forwards you a 15-second clip in the group chat. It shows the head teacher of your school appearing to say something racist at what looks like a private party. The clip is grainy, it's short, but it looks pretty real. Twenty minutes in, it's been forwarded to 40 people. Your friend asks, 'Have you seen this??'

You have about 30 seconds before someone else in the chat replies and the story hardens. What's the strongest first move?