A deepfake is a photo, video, or audio clip in which artificial intelligence has replaced, altered, or entirely fabricated a real person’s face, voice, or likeness. The word fuses “deep learning” — the branch of AI that powers the technique — with “fake.” The result looks, sounds, and moves like the real person, but never happened.

The term was coined in late 2017 when a Reddit user going by “deepfakes” began posting face-swapped videos — the first widely shared examples of the technology. Since then the tools have become far more powerful, shifting from generative adversarial networks (GANs) to diffusion models that produce markedly more convincing results.

How deepfakes are made

Most deepfakes fall into three categories: face-swap video (one person’s face replaces another’s in footage), voice cloning (a person’s voice is synthesized from a short audio sample), and synthetic images (faces that either don’t exist or replace a real person’s).

Earlier tools used GANs, where two neural networks compete — one generating fakes, the other trying to detect them — until the generator becomes convincing enough to fool the detector. Since 2023 the dominant approach has shifted to diffusion models, which start with random noise and progressively refine it into a realistic image. Diffusion models produce higher-quality results and have outpaced detection tools trained on older GAN artifacts, forcing researchers to rebuild their detectors from scratch.

What deepfakes are used for

Deepfakes have genuine legitimate uses. In film production they allow aging actors to appear younger, stunt doubles to wear a star’s face, or dialogue to be dubbed into another language while lip movements stay in sync. In healthcare, voice cloning has helped people with ALS preserve their own voice before losing the ability to speak naturally.

The harmful uses are more prominent. The most common is non-consensual intimate imagery — placing a real person’s face onto explicit content without their knowledge or consent. Other harms include fabricated videos of political figures designed to mislead voters, and financial fraud using a deepfake video call to impersonate an executive and authorize a wire transfer (a company in Hong Kong lost $25 million this way in 2024).

How to spot a deepfake

Detection is difficult, and no single sign is definitive — high-quality deepfakes are designed to pass casual visual inspection. That said, some patterns are worth watching for:

  • Lighting inconsistencies — the face is lit differently from the background, or shadows fall in the wrong direction.
  • Skin anomalies — skin looks unusually smooth or plastic, with no natural pore variation, especially near the hairline.
  • Unnatural eye movement — blinking that looks mechanical, or pupils that don’t reflect light consistently.
  • Shifting edges — the face or hair flickers, blurs, or warps when the subject moves.
  • Audio-video mismatch — lip movements don’t quite sync with the words spoken.

Several free tools can assist. Deepfake Detector provides 50 free scans per month for video, image, and audio. Reality Defender offers a comparable free tier. No tool is infallible — healthy skepticism toward media from unknown sources remains the most reliable defense.

In the news

The EU’s updated AI Act, approved in 2026, requires AI systems to label synthetic audio, images, and video in machine-readable format starting August 2026. From December 2026 it bans the creation of non-consensual intimate deepfakes across the bloc. Providers who fail to comply face fines of up to €15 million or 3% of global turnover. For the full picture of what changed, see EU Extends High-Risk AI Deadlines and Bans Deepfake Intimate Images in AI Act Overhaul and the evergreen guide What Is the EU AI Act — and Does It Apply to Georgian Companies?.

Sources: Deepfake — Wikipedia · EU AI Act, Article 50 · Regulation (EU) 2024/1689