Pre-Publication V2. A class of artificial intelligence models that train on learn the underlying statistical patterns, structures, and relationships within training data in order to synthesize novel content including text, images, video, audio, code, 3D geometry, and other digital artifacts that reflect those learned distributions. without directly reproducing training examples. Unlike discriminative AI models, which classify or predict from existing data, generative models produce new outputs by sampling from learned representations of the data space. Generative AI encompasses a range of architectures including diffusion models, generative adversarial networks (GANs), variational autoencoders (VAEs), and transformer-based models, and serves as the foundational technology underlying most current synthetic media, large language models, and multimodal AI systems.
Deliberation Summary:
(1) Recommendation not to describe the output as “novel”, or include that the models do not directly reproduce training examples, because there is published research demonstrating training material can be extracted from such models at scale, that the models are trained to reconstruct their training set, and that simple prompts can return close copies of existing material, including recognizable performers and on-screen text, without a user asking for them, noting that this behavior is inherent to the technology and is central to ongoing litigation and to the ethical and legal framing of creative uses. (2) Notation that the verb describing how a model acquires patterns carries weight in copyright litigation suggesting a neutral alternative.
Pre-Publication V1. A class of artificial intelligence models that learn the underlying statistical patterns, structures, and relationships within training data in order to synthesize novel content including text, images, video, audio, code, 3D geometry, and other digital artifacts that reflect those learned distributions without directly reproducing training examples. Unlike discriminative AI models, which classify or predict from existing data, generative models produce new outputs by sampling from learned representations of the data space. Generative AI encompasses a range of architectures including diffusion models, generative adversarial networks (GANs), variational autoencoders (VAEs), and transformer-based models, and serves as the foundational technology underlying most current synthetic media, large language models, and multimodal AI systems.