Risks & safety
Deepfake
Also known as: synthetic media
A fake but realistic image, video or audio clip made by AI, typically showing a real person saying or doing something they never did.
A deepfake is a realistic fake made by AI, a video, photo or voice clip that shows a real person doing or saying something that never happened. The technology can be used harmlessly (film effects, satire) but also for serious harm: scams, fake news, and fake images of real people. Because deepfakes keep getting more convincing, learning to be sceptical of "video proof" matters more than ever.
A deepfake is synthetic media, video, image or audio, generated or manipulated by AI to depict something fictitious as if real, most often a real person. It uses generative AI techniques such as diffusion and face/voice synthesis.
Where the harm is
- Fraud: cloned voices for scam calls, fake video of executives authorising payments.
- Disinformation: fabricated clips of public figures.
- Personal abuse: non-consensual imagery.
Defences
- Detection: tools that spot AI-generated artefacts (an ongoing arms race).
- Provenance: watermarking and content-credential standards that label AI media.
- Awareness and process: verifying identity through separate channels, not trusting media alone.
Deepfakes are AI-synthesised or manipulated media, historically autoencoder/GAN-based face swaps, now increasingly diffusion- and neural-voice-based, producing high-fidelity fabricated audio-visual content. The core societal problem is the erosion of media as evidence and the scalability of targeted fraud and disinformation, including a "liar's dividend" whereby genuine media can be dismissed as fake.
Countermeasures fall into passive detection (media forensics exploiting generation artefacts and inconsistencies; Verdoliva 2020) and active provenance (cryptographic content credentials such as C2PA, and generative watermarking). Detection is adversarially brittle and generalises poorly across generators, so provenance and process controls (out-of-band verification, authentication) are increasingly emphasised. Governance responses include disclosure/labelling requirements (e.g. transparency obligations for synthetic media in the EU AI Act) and platform policies. It sits within broader AI ethics and safety concerns around misuse of generative models.