# What is Deepfake?

Also called Synthetic Media Impersonation.

A deepfake is synthetic audio, image or video that convincingly depicts a real person saying or doing something they did not. The term covers a spectrum from consensual entertainment and dubbing to non consensual intimate imagery, political disinformation and fraud, and the harms and legal treatment differ sharply across that spectrum.

Generation quality has improved to the point where visual inspection is unreliable for a general audience, and voice cloning in particular can be convincing from short reference samples. This has shifted the practical defense away from detecting artifacts and toward verifying the channel: confirming identity through a separate trusted path rather than trusting the apparent voice or face in a single call or clip.

The most financially damaging pattern in organizations is impersonation of authority to authorize an action, such as a payment or a credential reset. Defenses are procedural rather than technical: require out of band confirmation for high value actions, use a shared code word for urgent requests, enforce dual authorization on payments, and train staff that urgency plus secrecy is itself the warning sign.

Automated detection exists but degrades quickly as generation methods change, and it produces both false positives and false negatives at rates that make it unsuitable as sole evidence. Provenance and watermarking approaches attack the problem from the other end by making authentic content verifiable, which is more durable than trying to identify every synthetic artifact after the fact.

Legal responses are developing unevenly. Some jurisdictions have criminalized non consensual intimate imagery and regulated synthetic political advertising, transparency obligations for synthetic content appear in the EU AI Act, and proposals addressing voice and likeness rights have been introduced in various legislatures. Platform policies typically prohibit deceptive synthetic depictions of real people and require labeling for altered media.

## Key points

- Synthetic media depicting real people in fabricated situations
- Visual and audio inspection is no longer a reliable defense
- Out of band verification is the practical control for impersonation
- Automated detection is unreliable as sole evidence
- Legal treatment varies and is still developing

## In practice

A finance employee receives a video call from someone who looks and sounds like an executive, urgently requesting a transfer to a new account and asking that it stay confidential. The company's rule requires any new payee to be confirmed by calling a number from the internal directory and to be approved by a second signer. The employee follows it, the executive knows nothing about the request, and no funds move.

## Related terms

- [Content Provenance](/en/glossary/content-provenance)
- [Watermarking](/en/glossary/watermarking)
- [Content Moderation](/en/glossary/content-moderation)
- [AI Incident Response](/en/glossary/ai-incident-response)

[Back to the AI Glossary](/en/glossary)