If generative AI is the answer, what is the question?

2025-09-07

AI Summary

This paper addresses a foundational question in generative AI: What is its intrinsic nature as a distinct machine learning task, and how can generative tasks be formally characterized and theoretically related to prediction, compression, and decision-making? To this end, the authors propose a task-centric research paradigm and develop a unified theoretical framework integrating probabilistic modeling and two-player game theory, rigorously distinguishing density estimation from sampling-based generation. The framework systematically unifies five major generative paradigms—autoregressive models, variational autoencoders, normalizing flows, generative adversarial networks, and diffusion models—and incorporates post-training alignment strategies while embedding socio-ethical considerations. The resulting formal foundation advances the theoretical understanding of generative AI and provides systematic support for responsible AI practices, including privacy-preserving generation, content provenance, and copyright-compliant deployment.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Philosophy and Ethics of AI: Artificial General Intelligence

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Beginning with text and images, generative AI has expanded to audio, video, computer code, and molecules. Yet, if generative AI is the answer, what is the question? We explore the foundations of generation as a distinct machine learning task with connections to prediction, compression, and decision-making. We survey five major generative model families: autoregressive models, variational autoencoders, normalizing flows, generative adversarial networks, and diffusion models. We then introduce a probabilistic framework that emphasizes the distinction between density estimation and generation. We review a game-theoretic framework with a two-player adversary-learner setup to study generation. We discuss post-training modifications that prepare generative models for deployment. We end by highlighting some important topics in socially responsible generation such as privacy, detection of AI-generated content, and copyright and IP. We adopt a task-first framing of generation, focusing on what generation is as a machine learning problem, rather than only on how models implement it.

Problem

• Exploring generative AI as a distinct machine learning task
• Surveying five major generative model families and frameworks
• Addressing socially responsible generation including privacy and copyright

Innovation

• Probabilistic framework distinguishing density estimation from generation
• Game-theoretic two-player adversary-learner setup for generation
• Post-training modifications preparing models for deployment