Textbook of AI

Deepfakes are AI-generated synthetic media, images, audio recordings, or video, that depict real people doing or saying things they did not. The term is a contraction of "deep learning" and "fake" and entered public usage in 2017 via a Reddit user posting face-swap pornography. Underlying technologies have advanced from per-clip fine-tuned models in 2017 to single-image zero-shot synthesis with diffusion-based and transformer-based models by 2024–2026.

Mechanisms

Three principal generation approaches:

• Face swap / face reenactment, replace one person's face with another's while preserving expression and head pose. Original autoencoder-based methods (DeepFaceLab, FaceSwap) gave way to GAN-based and now diffusion-based pipelines.

• Voice cloning, a few seconds of reference audio is sufficient for systems like ElevenLabs, OpenVoice and Tortoise TTS to produce arbitrary speech in the target's voice.

• Whole-clip generation, text-to-video models (Sora, Veo, Kling, Runway) synthesise people from scratch, including specified real individuals if conditioned on their likeness.

Harms

Documented categories:

• Non-consensual intimate imagery (NCII), by far the largest category by volume; overwhelmingly targets women.

• Election interference, fake clips of politicians (the 2024 US, Indian and EU election cycles all saw documented incidents).

• Fraud, voice-cloned CEOs authorising wire transfers; documented losses in the tens of millions.

• Defamation and harassment, fake confessions, fake compromising material.

Detection

Detection is a cat-and-mouse problem. Indicators that worked in 2019 (inconsistent eye reflection, frame-level artefacts, anomalous face landmarks) are largely defeated by 2024-era generators. Active detection research focuses on:

• Frequency-domain artefacts, diffusion models leave subtle spectral fingerprints.

• Physiological consistency, pulse signals from skin micro-colour change.

• Multimodal mismatch, audio-visual phoneme/viseme inconsistency.

• Provenance, falling back on watermarking and C2PA when model-internal cues fail.

Defences and policy

• Platform policy, most major platforms now prohibit non-consensual deepfakes; enforcement varies.

• Legislation, the UK 2023 Online Safety Act criminalised non-consensual deepfake distribution; the US Take It Down Act (2024) extended this; the EU AI Act mandates labelling.

• Provenance, C2PA Content Credentials and SynthID-style watermarking are the principal technical defences.

Status

As of 2026, deepfake quality is high enough that human verification by content alone is unreliable; defence has shifted to provenance over detection. The volume of deepfake content online doubles roughly every six months by some industry estimates.

References

• Korshunov, Marcel (2018). DeepFakes: a New Threat to Face Recognition?

• Citron, Chesney (2019). Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security.

• UK Online Safety Act (2023); EU AI Act (2024); US Take It Down Act (2024).