This is the second in our plain-English series about artificial intelligence. You can read Part 1 — What AI actually is — at news.atozofcyber.co.uk
Not long ago, spotting a fake image was fairly straightforward. The hands were wrong — too many fingers, fused knuckles, impossible joints. The text in the background was garbled. The lighting on a face did not match the room. These were the tells that gave AI-generated images away, and for a while they were reliable.
They are no longer reliable. The technology has moved on fast — much faster than most people realise. And that matters, because the ability to tell real from fake is one of the most important new skills of the digital age. Not just for journalists or fact-checkers. For everyone.
So why does this matter?
AI-generated content — images, video, audio, and text — is now everywhere. Most of it is harmless or even useful: illustrations for articles, voices for audiobooks, background music for videos. But some of it is designed to deceive. Fake images of real people saying or doing things they never did. Fake voices making phone calls. Fake videos used in scams, in political disinformation, and in the kind of social engineering attacks we covered on this site in March.
In January 2025, a UK teacher was driven into hiding after a deepfake video falsely showed her making racist remarks. In the same year, Europol documented individual financial fraud cases exceeding £20 million that were carried out using real-time deepfake video calls of fake executives authorising wire transfers. These are not edge cases. They are the leading edge of something that is growing.
The good news is that you do not need to become a forensics expert. You need to develop a habit of pausing, and to know what to look for.
Spotting AI-generated images
The old advice — look for extra fingers, strange hands, garbled text — is less reliable than it was. By 2025, the major image generation tools had largely solved those problems. But new tells have emerged in their place.
The edges and boundaries. Where hair meets background, where a collar meets a neck, where glasses meet a face — AI still struggles with these transitions. Look for slight softening, blurring, or an unnatural smoothness where two distinct things meet.
Texture consistency. AI images often have a strange uniformity to them. Skin looks slightly too smooth across the whole face. Fabric has a repeating quality. Background details that should be distinct — bricks, leaves, crowd members — can look oddly similar to each other.
The eyes. Reflections in eyes should match the environment the person is in. In AI-generated images, the reflections are often inconsistent between left and right, or show a light source that does not match the rest of the image.
Context that does not quite fit. A famous person in an unusual or implausible situation. An image that perfectly illustrates an outrage-inducing claim. A photograph that seems almost too perfectly composed for the moment it supposedly captures. These are not proof of fakery, but they are reasons to pause.
Most importantly: reverse image search. Before sharing any image that surprises or outrages you, drag it into Google Images or TinEye. If it appears in different contexts, on different sites, or the search turns up contradictory information, that is a signal worth heeding.
Spotting AI-generated video and audio
Video deepfakes have advanced dramatically. A decent home computer can now generate convincing deepfake video in real time. The old advice — watch for unnatural blinking, strange mouth movements, lighting mismatches — remains worth knowing, but the technology is improving faster than these detection methods.
What has not yet been fully solved is consistency over time. Watch for subtle shifts in skin tone between frames. Watch for the edges of the face against hair or background — they tend to shimmer or blur slightly in motion. Watch for moments when a person turns their head sharply or looks to one side — these transitions still catch AI systems out.
For audio, AI voice cloning has become frighteningly capable. A convincing clone of someone's voice can be generated from as little as three seconds of audio harvested from a social media video. If you receive a phone call from someone you know asking for something unusual — money, a password, urgent action — and something about the voice or the request feels slightly off, trust that instinct and call them back on a number you already have.
AI-generated text: the hardest to spot
Text generated by AI is, in many ways, the hardest AI content to detect reliably. The tools available to the public for detecting AI-written text are inconsistent — they flag human writing as AI and miss AI writing as human with enough frequency that they cannot be relied upon alone.
What you can look for is a certain quality in the writing itself. AI-generated text tends to be competent but slightly characterless. It is rarely wrong, but rarely surprising. It tends to hedge and qualify. It uses certain phrases — "it is worth noting," "it is important to understand," "in conclusion" — with a frequency that human writers naturally avoid after a while. It is fluent but somewhat frictionless, as if it was designed to not offend any reader.
None of these are definitive. But combined with other signals — a website you have not seen before, a claim you cannot verify elsewhere, a piece that seems designed to confirm exactly what you already believe — they are worth weighing.
The most practical thing you can do
The single most useful habit is the simplest one: pause before sharing.
The entire machinery of AI-generated disinformation depends on speed — on outrage or surprise or delight moving content through social networks before anyone stops to check. The pause is the intervention. Ask yourself: where did this come from? Can I verify it independently? Does it seem too perfect, too outrageous, or too convenient?
Verification tools exist and are worth knowing about. The Coalition for Content Provenance and Authenticity — C2PA — is a developing standard that embeds a cryptographic signature into authentic media, creating a verifiable chain of custody. An increasing number of cameras, platforms, and news organisations are beginning to support it. Content that carries a valid C2PA signature is verifiably authentic. Content that lacks one is not necessarily fake — but it cannot be verified in the same way.
Free tools like TrueMedia.org allow anyone to submit an image or video for analysis. They are not infallible, but they are useful.
What does this mean for me?
Pause before you share. If an image or video surprises, outrages, or delights you — that emotional reaction is exactly what people who create fake content are aiming for. Pause. Check. Then decide.
Reverse image search anything suspicious. It takes fifteen seconds and has caught thousands of fakes.
Be especially cautious with audio. Voice cloning is now trivial and very convincing. Any unexpected phone call requesting money, passwords, or urgent action should be verified by calling back on a number you already know.
Talk to children and older relatives about this. Both groups are targeted disproportionately by fake content — children because they are still developing critical thinking habits, older relatives because they may be less familiar with the technology. Showing them one example of a convincing fake is worth more than a long conversation about it.
Remember: being fooled is not a failure. These tools are designed by professionals to deceive. The goal is not perfection — it is developing the habit of asking the question.
🧠 The Human Factor
| Technology involved | AI image generation, video deepfakes, voice cloning, and AI-written text — all now accessible to anyone with a consumer device |
| Root cause | The same technology that makes AI creative tools useful also makes convincing fake content trivially easy to produce — and human psychology makes us share surprising or emotional content before checking it |
| What was at risk | Reputation, financial safety, political trust, and personal relationships — all of which have been demonstrably damaged by AI-generated fake content in real cases |
| Prevention | Pause before sharing; reverse image search; verify audio calls independently; use emerging tools like C2PA verification and TrueMedia.org |
Next in the AI series: AI agents — when AI stops talking and starts doing things.
References and sources
- Global Investigative Journalism Network: Reporter's Guide to Detecting AI-Generated Content (2026) — gijn.org
- Europol Internet Organised Crime Threat Assessment 2025 — europol.europa.eu
- Coalition for Content Provenance and Authenticity — c2pa.org
- TrueMedia detection tool — truemedia.org
- MIT Media Lab: Detect DeepFakes project — media.mit.edu
- UK case: deepfake video of teacher, reported January 2025 — multiple UK news sources