Generative AI
A class of AI systems that produce new content — text, images, audio, video, or code — rather than classifying or analyzing existing data. Includes diffusion models, large language models, GANs, and TTS systems. Deepfake generation is a subset.
Generative AI is the broad category that deepfake generation belongs to. It covers any AI system whose output is new content: new text (LLMs), new images (diffusion models, GANs), new audio (TTS, voice cloning), new video (text-to-video models), new code (code-generation models).
The two families
Most generative AI falls into one of two architectural families:
- Autoregressive. Generate the output one token at a time, each conditioned on what came before. Standard for text (GPT, Claude, Gemini, Llama) and some audio models.
- Iterative denoising. Start from noise and refine toward a coherent output over multiple steps. Standard for images (diffusion models) and increasingly audio and video.
GANs are a third, older family. Once dominant for image generation, now mostly displaced by diffusion outside of real-time applications.
Legitimate vs. deceptive use
Generative AI is not inherently a deepfake. The distinguishing feature of a deepfake is intent to deceive. AI-generated illustrations, localized voiceovers, coding assistants, and text summaries are all generative AI and not deepfakes.
The regulatory landscape is converging on "disclose if synthetic" rather than "prohibit synthetic" — see the EU AI Act and similar frameworks.