AIGC (Artificial Intelligence Generated Content) refers to the production method where AI models automatically generate content such as text, images, audio, video, and code. It contrasts with "discriminative" AI (classification, recognition, retrieval): the latter provides judgments, while the former produces finished products. It changes the starting point of content production—humans no longer work word by word or frame by frame, but instead provide instructions, examples, or reference materials, allowing models to generate initial drafts or even final drafts.

The technical origins of AIGC lie in generative models. In 2013, Kingma and Welling proposed Variational Autoencoders in the paper Auto-Encoding Variational Bayes, using latent variables to model data distributions. In 2014, Goodfellow et al. used adversarial training between generators and discriminators in the paper *Generative Adversarial Networks * to significantly enhance the realism of generated images. In 2020, Ho et al. proposed diffusion models in the paper *Denoising Diffusion Probabilistic Models *, replacing adversarial training with "step-by-step denoising," resulting in better sample quality and training stability.

The direct impact of AIGC is a significant reduction in the unit cost of content production, with design, marketing, film and television, education, and gaming industries restructuring their respective workflows. Simultaneously, it has brought several governance issues to the forefront: copyright and licensing of training data, the authenticity and traceability of generated content (watermarking, content credentials), and trust risks posed by deepfakes. It should be noted that the abbreviation "AIGC" is most widely used in Chinese industrial and policy contexts, whereas English literature more commonly uses "generative AI" or "AI-generated content," with both terms referring to essentially the same concept.