Weekinfo

Artificial intelligenceWeek 47

Deepfakes: how to spot an AI-generated image, video or voice, and what to do

Read in Français Español

Mains tenant un téléphone portable à l'écran flou, dans une pièce sombre éclairée par une lampe
Contents
  1. What clues still give away a fake image or video?
  2. How do you guard against a cloned voice on the phone?
  3. Which tools can help you check?
  4. What does the law say in the US, Europe and Spain?
  5. Victim or witness: what should you do?
  6. Frequently asked questions
  7. Sources

A fake video of a politician, a doctored photo shared after a disaster, a relative's voice begging for money: deepfakes are audio, photos or videos created or altered by AI, and they are now easy to make. The good news is that a few habits and free tools sharply cut the risk of being fooled, and the law got tougher in 2026 on both sides of the Atlantic.

What clues still give away a fake image or video?

France's data protection authority (CNIL) and Spain's cybersecurity institute (INCIBE) list common technical flaws. None is proof on its own, and newer generators fix many of them, but they should raise doubts.

  • Blurry or pixelated areas, especially around the face, hair and hands.
  • Unnatural eye movements or blinking.
  • A distorted mouth, or lips out of sync with the voice.
  • Inconsistent light and shadows, or a warped background.
  • Skin tone or face position that doesn't match the rest of the body.
  • A very short clip, or shocking content that spreads unusually fast.

The best habit is a reporter's one: who posted it first? Has a news outlet or an official account confirmed it? Who benefits from it spreading? Anything shocking deserves a check before you share.

How do you guard against a cloned voice on the phone?

Voice has become scammers' favorite tool. The Federal Trade Commission puts the rule bluntly: don't trust the voice. A clone can imitate a child, parent or boss from a few clips posted online.

  • Hang up and call the person back on a number you already know, never one given during the call.
  • If you can't reach them, contact another relative or a friend of theirs.
  • Be wary of any request for a wire transfer, cryptocurrency or gift cards.
  • Agree on a family question or code word that only your close circle knows.

These scams look a lot like text-message fraud: our guide to text and phone scams explains where to report them.

Which tools can help you check?

Two tracing systems are spreading. SynthID, from Google DeepMind, embeds an invisible watermark in generated content: you can check for it by uploading an image, video or audio clip to the Gemini app, or on the SynthID Detector portal, which also recognizes content from OpenAI, NVIDIA and Kakao. Content Credentials (the C2PA standard) work like an origin label you can read with the free tool at verify.contentauthenticity.org.

What does the law say in the US, Europe and Spain?

Main deepfake rules (November 2026)
WhereRuleSince
United StatesTAKE IT DOWN Act: platforms must remove nonconsensual intimate images, including AI-generated ones, within 48 hours of a valid requestFTC enforcement since May 19, 2026
European UnionThe AI Act requires deepfakes shown to the public to be clearly labeled; a code of practice offers shared iconsAugust 2, 2026 (code published June 10, 2026)
FranceCreating a montage of someone's image without consent: up to 1 year in prison and a €15,000 fine, more if sexual or posted onlineIn force
SpainDraft law on honor and image: commercial use of a person's voice or image via AI without consent deemed unlawfulApproved by the cabinet July 7, 2026, before parliament

For the European timeline in detail, see our explainer on the EU AI Act.

Victim or witness: what should you do?

  1. Don't share the content, even to warn people.
  2. Keep evidence: screenshots, links, messages.
  3. Report it to the platform hosting it.
  4. In the US, report scams at ReportFraud.ftc.gov, and platforms that fail to remove an intimate image at TakeItDown.ftc.gov.
  5. In France, report serious illegal content on Pharos (internet-signalement.gouv.fr) and file a police complaint.
  6. In Spain, call INCIBE's free 017 helpline, every day from 8 a.m. to 11 p.m.

Frequently asked questions

Can you be 100% sure an image is fake?

Rarely. Visual clues and detectors give signals, not certainty. Tracing the original source remains the most reliable method.

What must US platforms do about intimate deepfakes?

Under the TAKE IT DOWN Act, enforced by the FTC since May 19, 2026, covered platforms must offer a removal process and take down such images and known copies within 48 hours of a valid request.

Do platforms have to label AI content in Europe?

Since August 2, 2026, the EU AI Act requires deepfakes shown to the public to be labeled by whoever publishes them.

How can I protect elderly relatives from cloned voices?

Share the FTC rule: don't trust the voice, hang up and call back on a known number, and refuse any urgent payment.

Sources

Read next

Sections

Language