The word "deepfake" entered mainstream vocabulary around 2018, yet it is still frequently confused with other types of visual manipulation. Is an image generated by Midjourney a deepfake? What about a video where a politician appears to say things they never actually said? Let's set the record straight.
In 2026, the distinction matters more than ever: visual manipulation tools have multiplied, malicious use cases have diversified, and the ability to tell these different forms of manipulation apart has become a basic civic skill.
Deepfake vs AI image: what's the difference?
The confusion is common, but the two concepts are distinct.
An AI image (or generative image) is created from scratch by an algorithm based on a text description. It does not depict any real person or scene — it is an entirely synthetic creation. Midjourney, DALL-E, and Stable Diffusion produce this type of image.
A deepfake, on the other hand, starts from existing material — a photo or video of a real person — and manipulates it to make them say or do something they never said or did. Here, AI is used to map one person's features onto another body, or to alter their expressions and voice in a convincing way.
| Generative AI image | Deepfake | |
|---|---|---|
| Starting point | Text (prompt) | Real photo/video of a person |
| Person depicted | Fictional or composite | Real, manipulated |
| Primary risk | General misinformation | Reputation damage, fraud |
| Example tools | Midjourney, DALL-E, Stable Diffusion | DeepFaceLab, FaceSwap, AI video tools |
In practice, the two categories increasingly overlap. It is now possible to generate an entirely AI-created image that realistically depicts an existing celebrity — technically AI generation, but with the social impact of a deepfake.
How does a deepfake work?
The earliest deepfakes relied on GANs (Generative Adversarial Networks): two neural networks compete in a loop, one generating fakes and the other trying to detect them. This ongoing adversarial process gradually improves the quality of the output.
Today, diffusion models — the same architecture behind Midjourney and Stable Diffusion — are increasingly used for face-swapping and video manipulation. They produce more stable and consistent results than older GANs.
In concrete terms, a facial deepfake ("face-swap") requires:
- A set of images of the target subject (a few dozen to a few hundred are enough today)
- A source video onto which the face will be superimposed
- A model trained to map expressions and movements from one face to another
Voice deepfakes follow a similar logic: a model learns the timbre and intonation of a voice from recordings, then generates new words in that voice.
Tell-tale signs of a deepfake
Detecting a deepfake video is different from detecting a static AI image. Here are the specific clues to watch for:
Eye blinking
Early deepfake models struggled to replicate natural blinking. Although newer models have largely addressed this, blinking that is too infrequent, too frequent, or visually jerky remains a warning sign.
Facial contours
The transition line between the manipulated face and the neck or hair of the source person is often the hardest area to render naturally. A subtle halo, a "mask" effect, or a texture mismatch along the face boundary is one of the most reliable indicators.
Lighting inconsistencies
The lighting on the superimposed face does not always match the lighting of the scene. If the face appears to be lit differently from the rest of the body, or if cast shadows don't match the visible light source, something is off.
Mechanical movements and micro-expressions
Real human expressions involve dozens of muscles. A deepfake may miss micro-expressions — those fleeting twitches around the eyes or mouth lasting less than half a second that lend credibility to an expression.
Deepfakes in 2026: increasingly realistic
In 2018, an amateur deepfake could be spotted in seconds. In 2026, the best deepfakes go unnoticed by the general public and require forensic analysis tools to be detected with certainty.
Several factors explain this rapid evolution: the democratisation of tools (some are freely available online), the increase in available computing power, and the availability of large datasets of images and videos for model training.
The stakes are real: non-consensual pornographic deepfakes, political manipulation, financial fraud (fake CEO on a video call), and large-scale misinformation. France has strengthened its legal framework on the issue, but detection still largely falls to platforms and informed individuals.
How can you test your ability to spot deepfakes?
Theory helps, but nothing beats practice. Fake or Real features images that cover the full spectrum: from "classic" AI generations to the most advanced manipulations of 2025-2026 models. Every round is also a training session.
Can you tell the difference between a real photo and an AI-manipulated image? Test yourself now.
Play Fake or Real →With regular practice, you will develop a form of visual intuition that goes beyond the conscious clues listed in this article. It is this skill — recognising that an image "doesn't look right" — that makes all the difference when facing the most sophisticated deepfakes.
Summary
A deepfake manipulates the image of a real person, while an AI image creates something entirely from scratch. Both erode trust in visual media, but through different mechanisms and with different risks. The signs to watch for in a deepfake: facial contours, lighting consistency, eye blinking, and micro-expressions. And in every case, practice remains the best defence.