A fake made with AI that convincingly copies a real person's face or voice.
"Deepfake" blends "deep learning" and "fake". AI can put one person's face onto someone else's body in a clip, or copy a voice from a short recording, making it look as if someone said or did things they never did. Many are now so convincing that you can't tell just by looking.
The same tools can be used for film effects, but deepfakes also spread false news and power scams where someone pretends to be a family member. Faking a classmate's image without permission can do real harm and may be a crime. A growing number of places, including the EU, China and South Korea, have rules requiring AI-made content to be disclosed.
Think of it like this
Think of a friend who does a perfect impression of your teacher. Over the phone, you might be fooled. A deepfake is that kind of impression done by AI, for faces as well as voices, and far more convincingly.
For example
If a family member calls in a panic asking for money, hang up and call them back on the number you already know.
Common mix-ups
Mix-upIf you look carefully, you can always spot a deepfake.
ActuallyClues like odd blinking or strange-looking hands used to work, but such flaws disappear quickly as the technology improves. Rather than hunting for flaws on screen, it's more reliable to check who first posted it and whether trustworthy outlets report the same thing.
Mix-upEditing a friend's photo into something as a joke is fine.
ActuallyA face and voice belong to their owner, so even for a joke the basic rule is to ask first. In 2025 OpenAI also worked with the actors' union SAG-AFTRA to strengthen safeguards stopping Sora 2 from copying a real person's face and voice without permission.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- If an urgent call comes in a family member's voice, what should you check before acting?
- Activity
- Each group makes a "checking habits card". For three situations — a call, a message, a clip — write three steps such as "pause → check by a number or route you already know → tell an adult". Agreeing on a family question only you know is a good idea too (don't write the answer on the card). Then read the 2022 China (E0180) and 2026 EU (E1029) records and compare what each rule requires people to disclose.
Records about this
Source-checked moments from AI history.
Provenance labeling such as C2PA and SynthID moves from goodwill to a regulatory compliance requirement. Dec 2, 2026 is only a transition deadline for systems already on the market, not a postponement of Article 50 as a
A likeness-rights track takes shape, separate from the copyright track
Real-name authentication and labelling duties for deepfakes. The forerunner of generative AI regulation