Picture a perfect digital puppet: a deepfake is a synthetic media file (video, audio, or image), crafted by deep-learning algorithms, that makes a real person appear to say or do something they never actually did. The word itself merges deep learning and fake.
How it works
Most deepfakes rely on Generative Adversarial Networks (GANs) or diffusion models. Two neural networks compete:
- the generator produces forged images;
- the discriminator tries to spot the fake.
This rivalry is captured by a min-max game:
$$\min_G \max_D \; \mathbb{E}{x}[\log D(x)] + \mathbb{E}[\log(1 - D(G(z)))]$$
Through training, the generator becomes so skilled that the fake grows nearly impossible to detect with the naked eye.
Why it is a risk
Malicious uses are spreading: political disinformation, voice-cloning scams, financial fraud, and reputational harm (notably through non-consensual intimate content).
| Legitimate use | Abusive use |
|---|---|
| Film dubbing | Fake executive statements |
| Educational avatars | Voice-cloning scams |
| Heritage restoration | Electoral manipulation |
Fighting back
Defenses combine automated detection (spotting artifacts, blinking patterns, facial inconsistencies), digital watermarking, and provenance standards (such as C2PA) that certify a file's origin.
The deepfake embodies the paradox of generative AI: the very technology that produces stunning special effects can also erode our trust in what we see and hear.