Deepfake – UseCaseinAI

Q: How are deepfakes created?

Early deepfakes used GANs — two neural networks competing, one generating fake faces and one detecting fakes, until the generator produced images the detector could not distinguish from real. Modern deepfakes increasingly use diffusion models and face-swapping architectures that require less source footage and produce higher-quality results faster.

Q: What are the real harms of deepfakes?

Non-consensual intimate imagery (the majority of deepfakes online are non-consensual sexual content targeting real women), political disinformation (fabricated videos of politicians saying things they never said), financial fraud (voice cloning used to impersonate executives in phone calls authorising wire transfers), harassment, and erosion of trust in authentic video evidence.

Q: How can deepfakes be detected?

Detection tools look for subtle artefacts: unnatural blinking patterns, inconsistent lighting between face and background, blurring around hairlines and ears, audio-visual synchronisation errors, and statistical patterns in pixel distributions. However, detection is an arms race — as generation improves, detection must keep pace. No detection method is reliable against the most advanced deepfakes.

⚡ A deepfake is AI-generated video, audio, or image content where a real person appears to say or do something they never said or did. Created using deep learning, deepfakes range from harmless entertainment to serious harms — non-consensual intimate imagery, political disinformation, and voice-cloning fraud costing companies millions.

Category: Generative AI · Difficulty: Beginner · Last updated: 15 May 2026 · 5 min read

Deepfake — What It Is, How AI Creates Synthetic Media & Why It Is One of the Most Serious Risks of Generative AI

What is Deep Fake?

In 2019, a CEO of a UK energy company received a phone call from his parent company’s chief executive — or so he thought. The voice was familiar, the accent correct, the instructions clear: transfer €220,000 to a supplier immediately. He did. The voice was an AI-generated deepfake. The money was gone within an hour.

Deepfakes are synthetic media — video, audio, images — generated by AI to make a real person appear to say or do something they never did. The term emerged in 2017 from an online community that used deep learning to convincingly swap celebrity faces into other videos. The technology has since become accessible to anyone with a laptop and a few hours of source footage.

How Deep Fake are created ?

Source footage of the target person is collected — photos, video clips, audio recordings.
A deep learning model (historically a GAN, increasingly a diffusion model) trains on this footage to learn the person’s appearance and voice characteristics.
For face-swapping: the model replaces the face in target video with the learned face, matching lighting, angles, and expressions frame by frame.
For voice cloning: the model learns the target’s vocal characteristics and can generate new audio of them saying anything typed as text.
Post-processing smooths artefacts to make the result more convincing.
Tools like FaceSwap, DeepFaceLab, and commercial voice-cloning APIs have made this accessible to non-experts.

Real-world examples

A Hong Kong finance worker was tricked by deepfake video conference call into transferring HK$200 million (approximately $25 million USD) in 2024 — colleagues’ faces and voices were all AI-generated.
Deepfake audio of Slovak opposition leader Michal Šimečka was circulated two days before the 2023 Slovak election, appearing to show him planning to buy votes — timed to maximise damage before the story could be debunked.
Over 95% of deepfakes online are non-consensual sexual content, predominantly targeting women who are public figures — one of the most significant harms of the technology in practice.

Common pitfalls (for society, not practitioners)

Liar’s dividend — deepfakes do not just enable fake videos. They give anyone accused of genuine wrongdoing a plausible defence: “that video is a deepfake.” The mere existence of deepfake technology erodes trust in authentic video evidence.
Detection is an arms race — every improvement in detection spurs improvement in generation. Current detection tools catch yesterday’s deepfakes, not tomorrow’s.
Legislation is lagging — most jurisdictions lack specific deepfake laws. Non-consensual intimate imagery laws, fraud statutes, and election interference laws are being applied patchwork.
Accessibility is increasing — what required specialist ML skills in 2019 requires only a smartphone app in 2026. The barrier to creating convincing deepfakes has dropped dramatically.

Frequently asked questions

QUESTION 1 What is a deepfake in simple terms?

ANSWER 1 AI-generated video, audio, or images where someone appears to say or do something they never did. Created from source footage of the real person using deep learning.

QUESTION 2 How are deepfakes created?

ANSWER 2 GANs or diffusion models learn a person’s appearance and voice from source footage, then generate new synthetic content that looks and sounds like them saying or doing anything.

QUESTION 3 What are the real harms of deepfakes?

ANSWER 3 non-consensual intimate imagery, political disinformation, financial fraud via voice cloning, harassment, and erosion of trust in authentic video evidence.

QUESTION 4 How can deepfakes be detected?

ANSWER 4 Unnatural blinking, lighting inconsistencies, blurring around hairlines, audio-visual sync errors, and pixel-level statistical patterns. Detection is an arms race — no method is reliable against the most advanced deepfakes.

📬 Get one concept + one use case every Tuesday. Join the newsletter →