In depth
Spotting a deepfake: eye, ear and the law
AI-made images, voices and videos: the clues that still work, the habits that hold up, and what European and French law actually say.
Reviewed on August 22, 2026
What is a deepfake?
The word blends "deep learning" and "fake"; it surfaced on a Reddit forum in 2017. It covers three families: images (a face swapped or invented outright), voice (a clone that says whatever it is told) and video (lips re-synced to new audio, or a face replaced live). The principle is always the same: a model trained on millions of faces or voices learns to produce new ones, then is steered towards one specific person. What changed after 2023 is the cost: a few seconds of audio or a single photo is enough, with consumer tools, sometimes free.
Read the discovery version
Imagine a magic mask: you lay it over a video, and your cousin's face turns into your teacher's face, moving mouth included. That is a deepfake: a picture, a voice or a video made by a computer to make you believe a real person said or did something they never did. Sometimes it is for fun, like in the movies. Sometimes it is to trick people, and that is when you need to know how to spot it.
Image tell-tales, and why they fade
The classic clues still exist: hands and fingers, teeth too regular, strands of hair merging into the background, asymmetric glasses or jewellery, unreadable text and logos, inconsistent shadows and reflections, background geometry that does not hold, skin smooth as plastic. On a face, the reflections in both eyes should tell the same scene. But each model version fixes the flaws of the previous one, and the studies are harsh: against recent images, the human eye is barely above chance. One more clue, outside the picture: a "candid" photo that is too well framed, or that triggers exactly the expected emotion, deserves doubt.
Read the discovery version
It is like spot-the-difference. Look at the hands: six fingers, or fingers crossing in a strange way. Look for glasses melting into skin, earrings that differ left and right, writing on a sign that means nothing, a shadow falling the wrong way. But be careful: computers learn fast and make fewer mistakes every year. So the safest clue is not inside the picture: it is asking where the picture comes from.
Voice and video: the clone and the fake call
A voice clone is built from a few seconds of recording, and sounds all the more right the more samples it was given. The remaining clues: an intonation that is too flat, no breathing, oddly placed pauses, names or numbers mispronounced, a slight delay before each answer, silence that is too clean. On video, watch the mouth (the sounds "m", "b", "p" need closed lips), the edge of the face flickering when the head turns, the glasses and hair. On a live call, ask the person to turn to profile, to pass a hand in front of their face, or ask a question only the real person could know. The scam script itself never changes: urgency, secrecy, money. In 2024 an employee of the engineering firm Arup in Hong Kong wired about 25 million dollars after a video meeting in which the finance director and several colleagues were all .
Read the discovery version
With a few seconds of your voice taken from a holiday video, a computer can make "your" voice say anything. Thieves use it to call a grandma: "Grandma, it's me, I'm in big trouble, send money quickly!" The trick clever families use: a secret password only the family knows. If the voice on the phone does not know it, you hang up and call back the real number.
The habits that help you check
Source first: who is posting, how old the account is, where the original post lives. Corroboration next: are other outlets or fact-checking services (AFP Fact Check, Reuters Fact Check, Full Fact) covering it? Then the tools: reverse image search (Google Lens, TinEye, Bing), keyframe extraction from a video (InVID-WeVerify), and context checks (date, weather, language on signs, buildings). A new habit joins the list: look for "Content Credentials" (the C2PA standard), a kind of signed label that says which device or software created and edited the file, readable at contentcredentials.org/verify; Google likewise offers a detector for its invisible SynthID mark. Two limits to keep in mind: "AI detectors" return a probability, not a verdict; and the absence of a label does not prove a piece of content is genuine, any more than its presence proves it is true. When in doubt, do not share, and report.
Read the discovery version
When a picture looks strange or too incredible, play detective. One: who posted it? Two: are serious news outlets showing it too? Three: ask an adult to do a "reverse image search": you give the picture to a search engine, and it finds where it first appeared. And the golden rule: if it scares you or makes you very angry right away, breathe before you share.
What the law says, and what to do
The EU AI Act (Article 50) has applied since 2 August 2026: providers of generative AI must mark their outputs in a machine-readable way, and anyone publishing a in a professional setting must clearly state that it is artificial, with a lighter duty for evidently artistic or satirical works; fines can reach 15 million euros or 3% of worldwide turnover. On 10 June 2026 the Commission published a voluntary code of practice, with a common icon. In France, the SREN law of 21 May 2024 updated the Criminal Code: spreading a montage made with a person's image or words without consent, when it is not obvious or not expressly flagged, carries one year in prison and a 15,000 euro fine (two years and 45,000 euros online, Article 226-8); for a sexual deepfake, it is two years and 60,000 euros, three years and 75,000 euros online (Article 226-8-1). Under the Digital Services Act, very large platforms must make deepfakes recognisable and let users flag them. In practice: keep the evidence (URL, screenshots, dates), report on the platform, then to the police portal PHAROS; turn to the regulator Arcom if the platform does not act, and file a complaint; cybermalveillance.gouv.fr guides victims.
Read the discovery version
In Europe, since August 2026, a rule says: if you make a fake video of a real person with AI and show it, you must say so clearly, like a label reading "made by AI". In France, spreading a fake picture of someone without warning that it is a montage is forbidden; and if the picture is sexual, the punishment is much heavier. If you come across a fake that hurts someone, you do not share it: you show an adult, and it gets reported.