Article · 8 min read

What is a deepfake?

Published September 2026

A deepfake is a video, photo, or audio recording that looks and sounds real but has been created or manipulated using artificial intelligence. The underlying technology can replace faces, manipulate facial expressions, synthesise faces, and synthesise speech. The word comes from "deep learning" (the type of AI involved) combined with "fake". So deepfake just means: a fake made with deep learning.

The short version of why they matter: these tools are used most often to depict people saying or doing something they never said or did. That covers everything from a scam video of a celebrity promoting a dodgy investment to a fabricated video call that tricks a company into wiring millions of pounds to a criminal.

How do deepfakes actually work?

The original technique behind most deepfakes is something called a GAN, which stands for Generative Adversarial Network. The name sounds complicated, but the idea is straightforward. You have two AI programs running against each other: one tries to generate a fake that looks real, and the other tries to catch it out. The generator and discriminator algorithms analyse data from media samples, and the generator creates or manipulates media to collectively resemble those samples as nearly as it can. Over millions of rounds of this competition, the forgeries get very, very good.

For face-swapping specifically, a neural network-based deepfake autoencoder analyses source video to understand relevant attributes of the target, such as facial expressions and body language, then imposes these characteristics onto the original video using an encoder (which captures the attributes) and a decoder (which applies them).

Voice deepfakes work on the same principle but with sound. A GAN clones the audio of a person's voice, creates a model based on their vocal patterns, and uses that AI model to make the voice say anything the creator wants. Alarmingly, it takes as little as three seconds of audio to clone a voice with 85% accuracy.

More recently, a newer class of AI called diffusion models (the same family that powers tools like Midjourney and DALL-E) has largely replaced GANs for image and video work. The shift from GANs to diffusion models has resolved the training instability and visual artefacts that previously served as forensic indicators, enabling one-shot face animation and real-time voice cloning on consumer hardware. In plain language: the old telltale glitches are mostly gone, and you no longer need a supercomputer to do this.

There's more where this came from. New articles most weeks.Browse all articles →

Why are deepfakes so hard to spot?

There are a few reasons they fool people so effectively.

Familiar faces trigger trust. Familiarity bias makes people fill in the blanks when they hear or see a voice or face they know. If something looks like your boss or sounds like a famous politician, your brain assumes it is.
Lip sync and expressions are now near-perfect. Lip sync and facial micro-expressions are increasingly accurate, removing the clues that used to give fakes away.
Real-time delivery cuts off verification time. Real-time delivery takes away the window that post-production analysis would otherwise need. When you're on a live video call, you can't pause and run forensic checks.
Most people are already bad at this. Humans detect fake videos less than 25% of the time. And when researchers tested people in a 2025 study and explicitly told them to watch for synthetic content, participants were still 36% less likely to correctly identify a fake video than a fake image.

Who gets targeted, and what happens to them?

There are three main groups who tend to end up in deepfakes against their will.

Ordinary people, mostly women

The largest category of harmful deepfakes by volume is non-consensual intimate imagery: real people's faces placed onto explicit videos they have nothing to do with. This is not a fringe problem. The overwhelming majority of deepfake content online falls into this category, and the victims are almost always women. The psychological harm is severe and, in most countries until very recently, the law offered little protection.

Celebrities and public figures

Celebrities were targeted 47 times in Q1 2025 alone, an 81% jump over the total for all of 2024, while politicians were impersonated 56 times in the same period. A celebrity deepfake can leverage the person's social influence to build support for causes they do not endorse, propagate hoaxes, or damage their reputation by making it seem as if they are saying or doing things out of character.

Businesses and their employees

This is where the financial damage is most measurable. The Arup Hong Kong incident in 2024 set a grim benchmark: $25.6 million lost from a single deepfake video call. The attack involved multiple AI-generated participants deceiving a finance team member across 15 separate transactions. That is not an isolated case. Deepfake fraud losses in the US alone reached $1.1 billion in 2025, tripling from $360 million the year before.

The scale of the problem is growing fast. Roughly 500,000 video and voice deepfakes were shared on social media globally in 2023, a figure estimated to have reached 8 million by 2025. Nearly half of businesses globally now report being targeted by audio or video deepfake fraud.

There's a second problem: the liar's dividend

Beyond the harm caused by fake content, there is a quieter damage happening in the opposite direction. The existence of increasingly convincing deepfakes creates a "liar's dividend": authentic evidence can be dismissed as AI-generated, providing plausible deniability to those accused of genuine misconduct. A politician caught saying something embarrassing can now claim the video is fake. A criminal caught on camera can raise doubt. The technology does not just let people create lies; it also gives them a tool to deny the truth.

What is being done about it?

Laws

The US passed the TAKE IT DOWN Act in May 2025, mandating that platforms remove non-consensual intimate deepfake content. The UK's Online Safety Act covers similar ground. These are meaningful steps, though enforcement across borders remains patchy.

Detection tools

Deepfake detection tools are improving and new systems evaluate mismatches in light, shadows, and audio consistency. Some firms, like Reality Defender, aim to tag manipulated content before it can spread. The catch is that the best AI detection tools, while reaching 96% accuracy in controlled lab settings, drop 45 to 50% when deployed in real-world conditions. Detection is still losing the race against generation.

Watermarking and metadata

Detection tools are increasingly being augmented with watermarking, forensic tracing, and verified metadata flags that travel with content and signal whether it was AI-generated. This is part of why several countries (including the UK) are now requiring AI companies to watermark synthetic content. The idea is that even if you can't spot the fake by looking, the file carries a hidden label that detection software can read.

What can you do right now?

Most people are not going to be targeted by a $25 million video-call heist. But the more realistic threats, such as a voice deepfake of a family member claiming to be in trouble, or a fake video used to damage someone's reputation, are genuinely worth being aware of. A few practical habits help:

Verify through a different channel. If you receive a surprising or urgent request via video call or voicemail, call the person back on a number you already know. Do not use the contact details provided in the suspicious message.
Agree on a code word. Some companies are now using "deepfake passwords": phrases or gestures that executives agree on ahead of time and use on live calls to prove their identity. Families can do the same thing.
Slow down on urgent requests. Deepfake scams almost always manufacture urgency. The pressure to act fast is the attack.
Be sceptical of viral video, especially of public figures. Before sharing a shocking clip of a politician or celebrity, ask yourself whether you can verify it through a news outlet or the person's own official channels.

The technology is not going away, and it is going to keep improving. The best defence, for individuals and for organisations, is understanding how it works well enough not to be caught off guard by it.

From Telltale

Keep reading

If this one was useful, there's plenty more on the site. Pieces on how AI works, plus coverage of AI news, the downsides included. All free to read, no account needed.

See all articles →

References

Published September 2026 · telltale-ai.com
All articles · Privacy · Terms