Deepfakes 2026: protect yourself from AI scams (video, voice)

In one sentence

A deepfake is a video, audio or image created by AI that shows a person saying or doing something they never said or did. In 2026, it's become widespread and mainstream : and it can cost you dearly if you don't know how to spot them.

🎭

The analogy that works

A deepfake is like a brilliant actor who could perfectly imitate your boss on a video call: same voice, same gestures, same expressions. Except it was created in 30 seconds by an AI. And it's asking you to transfer £25 million to an account. Not so funny now, is it?

🎬 Want to understand how AI creates videos?

We explain how Veo 3, Sora and video generation tools work in 2026.

Read the AI video article

The case that panicked the whole world

January 2024, Hong Kong: an employee of a British multinational receives an urgent video conference invitation. On screen, his CFO and several colleagues ask him to approve a transfer of $25 million to "strategic" accounts.

Everything sounds right. The faces, the voices, the gestures. He approves.

Verdict: everyone on screen was a deepfake. The scammers collected public images (LinkedIn, conferences) and reconstructed an entire video conference in real time.

→ $25 million lost. It's the most expensive deepfake attack in history (to date).

And this is just the beginning

According to a recent study, more than half of European cybersecurity professionals believe that deepfake attacks will be the main scam vector in 2026. The tools have become accessible: for £50/month, anyone can generate a convincing deepfake video.

How does a deepfake work?

You don't need to be an engineer to understand:

The 4 steps to create a deepfake

1.
Collect images/audio of the target
10-30 seconds of voice + a few photos are enough. Often collected from LinkedIn, YouTube, social media.
2.
Train the AI model
An AI learns to imitate the person's voice, face, expressions (15 minutes to a few hours).
3.
Generate the content
The AI produces a video/audio where the person 'says' or 'does' what you want. Real-time possible in 2026.
4.
Distribution
Fake video call, false phone call, video posted online...

The 4 types of deepfakes in 2026

Deepfake variations

	🎭Type	💰Cost / Effort
Video (face swap)	Replacing one face with another in an existing video	Free (FaceFusion, DeepFaceLab)
Video (full generation)	Generating an entire video from text	£20-50/month (Veo 3, Sora)
Audio (voice cloning)	Cloning a voice from 10-30 seconds	Free (ElevenLabs free, Resemble.ai)
Image (face generation)	Ultra-realistic photos of non-existent or real people	Free (Midjourney, Stable Diffusion)
Live deepfake (real-time)	Live video call where your face is replaced in real time	Complex setup but accessible

The most common deepfake scams

📚6 scenarios multiplying in 2026

1. The "fake boss" scam

You receive a call/video call from your superior asking for an urgent transfer or confidential information. Their voice is perfect. Their face too. Except it's a deepfake.

How to protect yourself: golden rule = any unusual financial request must be validated via a 2nd channel (direct phone call, in person, Signal message). Never by email/video call alone.

2. The "your children in distress" scam

Phone call. Your child's voice crying: "Mum, I've had an accident, I need money right away". The scammers cloned their voice from 3 seconds of one of their Instagram Reels.

How to protect yourself: set up a family code word. If your child doesn't give it, it's a fake. No panic = no immediate payment.

3. The fake celebrity recommending an investment

Video of Elon Musk, Bernard Arnault or another celebrity recommending a miracle crypto investment. Everything is fake but it pushes you to invest.

How to protect yourself: no celebrity sends you personal recommendations. If it's too good to be true, it's a scam.

4. Political deepfake

"Leaked" video of a politician saying horrible things before elections. Massive distribution in a few hours. When the denial arrives, the damage is done.

How to protect yourself: don't share a shocking political video without checking 2-3 reliable sources. AFP Factuel, Les Décodeurs (Le Monde) do this work.

5. Revenge porn (intimate deepfake)

Fake intimate photos/videos of a person (often ex-partner, colleague) distributed to harass them. It's illegal in France (up to 2 years in prison).

How to protect yourself: immediate reporting to the platform + Pharos + file a complaint. Loi Avia + new 2024-2026 laws protect.

6. Banking identity theft

Your bank's "identity verification" service asks you for a selfie video. You give it. Later, your visual identity is used to validate fraudulent accounts.

How to protect yourself: NEVER make a selfie video by sending documents via email/SMS. Real banks use their secure app.

How to detect a deepfake?

The tools are not perfect but here are the signals:

Visual signals

👀 Gaze and blinking Deepfakes often have weird blinking (too fast, too slow, irregular). The gaze may seem empty or misaligned. 👄 Mouth and synchronisation On amateur deepfakes, the lips don't perfectly follow the words. On professional ones, harder to see. 💡 Light and shadows Check if the facial lighting matches the environment. Often a face that's too "smooth" or inconsistent shadows. 🦷 Teeth and ears AIs struggle with fine details: asymmetric teeth, weird ears, blurred contours. 📐 Face border Look for a blurred line or artifacts around the neck, hair, jaw.

Contextual signals (often more effective)

⚡ Abnormal urgency "Transfer the money in 10 minutes or..." → major red flag. Urgency is technique #1 of scams. 📞 Refusal to change channel "No, we can't call back directly, just this video call..." → suspicious. A real person accepts a callback. 🎯 Unusual request Your boss asking you for a transfer to a weird account, your mother not recognising a family reference... → verify. 🔐 Refusal of code word You ask for the family code word, the other evades or gets angry. → it's a fake.

Detection tools in 2026

Deepfake detector accuracy (2026)

AI watermarks (C2PA, ongoing)95% accuracy

Professional detectors (Reality Defender, Sensity)85% accuracy

Trained human eye70% accuracy

Consumer detectors (Microsoft Video Authenticator)65% accuracy

Untrained human eye50% accuracy

Reading: even the best tools are wrong 15% of the time. An unaware human is wrong 1 time in 2. The most promising technique: watermarks that AIs add automatically (C2PA standard).

Your anti-deepfake routine in 7 rules

🛡️

The 2026 survival kit

1. Family code word Choose a word known to all (not on social media). When in doubt → "What's the word?". No word = it's a fake. 2. 2nd channel for any financial request No transfer, no voucher, no gift card without confirmation by direct phone (remember, not the number that called you). 3. Beware of urgency "Right now", "before 6pm", "don't tell anyone" = scam signals. Take 5 minutes to verify. 4. Limit your public exposure The fewer videos of yourself you put on social media, the less AI can clone you. Private = safer. 5. Educate your loved ones 80% of scams target elderly people or those less aware. Explain to them. Do an exercise together. 6. Verify before sharing A shocking video of a public figure? Before retweeting: AFP Factuel, Les Décodeurs, Snopes. 7. Report Seen an illegal deepfake? Pharos (government) + platform where it's posted.

The legal framework in France (2026)

What the law says

In France, creating or distributing a deepfake without consent is punished: - Deepfake Law (June 2024): up to 3 years in prison and €75,000 fine for sexual deepfakes - Penal Code (art. 226-8): identity theft = 1 year in prison and €15,000 fine - European AI Act: mandatory transparency = all deepfakes must be labelled "AI-generated" since 2026 If you're a victim: 1. Screenshot/video (evidence) 2. Report to the platform (quick removal) 3. Complaint to police/gendarmerie 4. Pharos (https://www.internet-signalement.gouv.fr)

The metaphor that sums it all up

🎪

The illusionist in the AI era

Before, to make a convincing fake you had to be an exceptional magician: make-up artist, ventriloquist, actor. Very few people had the means. Today, anyone with £50/month and a PC can make a deepfake more convincing than a professional illusionist from 1980. → What's changed: it's not the deception technique, it's its democratisation. And what protected you before ("only a pro could do that") no longer works. You need to adopt new reflexes like we learned to be wary of phishing emails in the 2000s.

Key takeaways

✅ Deepfakes are massive in 2026: $25M lost by a UK company in 1 video call
✅ Tools accessible for £50/month or free
✅ Human detection = 50% accuracy only
✅ The 3 golden rules: family code word, 2nd channel, suspicion of urgency
✅ Educate your loved ones: they're the ones targeted most
✅ Report: French law protects, but you must act fast

Deepfakes are not going away. But with the right reflexes, you can protect yourself and your family.

🧠 Quiz

Question 1 of 3

How much voice time is needed to clone someone in 2026?

Going further

🎬 Veo 3, Sora: creating AI videos, how these tools work
🌀 Understanding AI hallucinations, another major AI risk
🔒 AI and privacy

Deepfakes: how to protect yourself in 2026