A video of a public figure says something shocking. A voice note that sounds exactly like your boss asks you to make an urgent payment. A photo shows an event that never happened. A few years ago you could mostly trust your own eyes and ears online. Today, you can't always — because deepfakes have made convincing fakery cheap and easy. Understanding what they are and how to spot them has quietly become an essential modern skill. Here's what you need to know.
What it is
A deepfake is media — video, image or audio — that has been created or altered by artificial intelligence to convincingly show someone saying or doing something they never actually did. The word stitches together "deep learning" (the AI technique behind it) and "fake." The same family of tools can swap one person's face onto another's body, put new words in someone's mouth, or clone a voice so accurately that close friends are fooled.
What makes deepfakes different from old-fashioned photo editing is scale and ease. Manipulating media used to take skill, time and money. Now AI does the heavy lifting, learning a person's appearance or voice from existing footage and generating brand-new, synthetic content that mimics them. The result is fakery that's faster to make, harder to detect, and available to almost anyone.
How deepfakes are made
You don't need to be technical to grasp the idea. An AI system is trained on lots of examples — many images and videos of a face, or recordings of a voice. By processing all that material, it learns the patterns: how the face moves when someone smiles or speaks, the rhythm and tone of a particular voice. Once trained, it can generate new media that follows those patterns, producing a face or voice that wasn't really there.
Two developments have made this genuinely alarming:
- Less source material is needed. Early deepfakes needed hours of footage. Some modern voice-cloning tools can imitate a person from only a few seconds of audio — easily scraped from a voicemail greeting or a social media clip.
- The tools are accessible. What once required expertise and powerful computers is now packaged into apps and online services. That democratisation is the real story: the barrier to creating a convincing fake has collapsed.
This is one of the more sobering sides of modern technology, and it's why deepfakes have become a mainstream cybersecurity concern — the same advances that power helpful tools also power convincing impersonation.
The real risks
It's tempting to think of deepfakes as celebrity hoaxes or internet curiosities. The more serious harms are closer to home and more practical:
- Fraud and scams. This is the big one. Criminals clone a familiar voice — a relative, a manager, a company executive — to authorise payments or extract information. A panicked "it's me, I'm in trouble, please send money" hits far harder in a loved one's actual voice.
- Misinformation. Fake videos of public figures can spread false claims fast, especially around elections or crises, eroding trust in what people see.
- Impersonation and reputational harm. Fabricated clips can make someone appear to say or do damaging things, with real consequences for individuals and organisations.
- Non-consensual content. Among the most harmful uses is placing people's likenesses into explicit material without consent — a serious abuse that several jurisdictions, including the UK, have moved to criminalise.
The unsettling shift deepfakes create isn't just "fake things look real." It's the flip side: real things can now be dismissed as fake. When anything could be a deepfake, bad actors can wave away genuine evidence as "probably AI." That erosion of shared trust is the deeper danger.
How to spot the visual and audio signs
The technology keeps improving, so no checklist is foolproof — but glitches still appear, particularly in lower-effort fakes. It's worth knowing what to look and listen for.
| In video and images | In audio |
|---|---|
| Unnatural or rare blinking | Flat, emotionless delivery |
| Odd lighting or skin texture | Strange pacing or pauses |
| Blurring/warping around the face edges | Mismatched background noise |
| Lip movements out of sync with words | Slight robotic or metallic tone |
| Hands or ears that look "off" | Breathing that sounds wrong or absent |
Look especially at the boundaries — where a face meets hair, neck or background — and at fine details like teeth, eyes and jewellery, which AI often renders imperfectly. For audio, genuine human speech has natural emotion and imperfections; a clone can sound subtly too smooth or oddly paced.
That said, be honest with yourself: the best deepfakes now pass these tests. So treat visual and audio cues as helpful hints, not a reliable verdict.
The strongest defence is context
Because the media itself is getting harder to judge, the most reliable protection has shifted from examining the pixels to questioning the situation. This is the same instinct that protects you from other digital deception, from phishing emails to scam websites: slow down and interrogate the source and the ask.
A few practical habits:
- Question the source and channel. Did this come from an official, verified account, or a random forward? Is a reputable news outlet reporting the same thing? A shocking "clip" that only exists on obscure accounts deserves deep scepticism.
- Be suspicious of urgency. Scams using cloned voices rely on panic and pressure — "do this right now, don't tell anyone." Urgency plus an unusual request is a classic red flag, exactly as it is with impersonation scams generally.
- Verify through a separate channel. If "your boss" or "your relative" makes an unexpected request, contact them another way — call the number you already have. A real person will be glad you checked; a scammer is stopped cold.
- Agree a safe word. Some families and teams now set a private phrase to confirm identity in an emergency call. It's a simple, powerful counter to voice cloning.
Where this is heading
It's worth being clear-eyed. Detection tools exist, and platforms are working on labelling AI-generated content, but it's an arms race — as detection improves, so do the fakes. Regulators such as Ofcom and bodies like the NCSC are increasingly focused on synthetic media, and the law continues to adapt to abusive uses. For now, though, technology won't fully solve this for us. Personal awareness — knowing deepfakes exist, recognising the signs, and verifying before acting — remains the front line.
The bottom line
A deepfake is AI-generated or AI-altered media designed to convincingly impersonate someone, and the technology has become realistic, cheap and widely available — including voice clones built from mere seconds of audio. The real dangers are practical: fraud, misinformation and impersonation, not just viral hoaxes. Visual and audio glitches can still give fakes away, but as they improve, your best defence is context: question the source, distrust urgency, and verify unexpected requests through a separate, trusted channel. In a world where seeing is no longer believing, a moment's healthy scepticism is the most reliable tool you have.