Deepfake Technology

A deepfake is synthetic or altered audio, video or imagery made by artificial intelligence to convincingly imitate a real person or event. Using deep learning, the system learns patterns in a person’s face, movements or voice and then generates new content that can appear authentic, although the depicted action or speech never occurred.

How a deepfake is created

Unlike conventional editing, which alters content manually, deepfake systems learn statistical patterns from examples and reproduce them automatically.

• The system is trained on photographs, video or audio of the target and learns features such as facial landmarks, expressions, speaking style, pitch and rhythm.
• It may perform face-swapping, facial reenactment, lip-synchronisation or voice cloning, depending on the intended output.
• Generated material is composited into the original recording, with lighting, colour, timing and sound adjusted to improve consistency.
• Repeated training and post-processing reduce visible or audible errors, making human detection more difficult.

AI models involved

Deepfakes can be produced through several generative AI techniques rather than a single fixed method.

• An encoder-decoder model compresses facial or vocal features into a learned representation and reconstructs them in the target form.
• In a Generative Adversarial Network, a generator creates synthetic samples while a discriminator tries to distinguish generated samples from genuine ones; their repeated competition improves realism.
• Diffusion-based models learn to generate content by reversing a gradual noising process and can create or modify highly detailed images and video.
• Speech-synthesis models learn the relationship between text, pronunciation and vocal characteristics to generate speech resembling a chosen speaker.

Why the output can deceive

The illusion becomes persuasive when identity, movement and sound remain mutually consistent. The technology itself can support legitimate creative or accessibility uses, but deception occurs when synthetic content is falsely presented as an authentic record.

• Identity mimicry can enable impersonation, fraud and unauthorised representation.
• Fabricated speeches or events can spread disinformation, damage trust and aggravate social tensions.
• Detection may examine artefacts, metadata and inconsistencies, while content provenance and authentication can help establish origin.

How UPSC asks this