What Is Generative AI?

Generative AI, also called generative artificial intelligence, is the branch of artificial intelligence concerned with models that learn the statistical structure of a training distribution well enough to produce new samples from it, whether those samples are text, images, audio, video, source code, or molecular structures. The output is called AI-generated content. The distinction from discriminative machine learning is what the model estimates: a classifier learns the probability of a label given an input, while a generative model learns the distribution of the data itself and can be sampled to yield artifacts that never appeared in training.

The field rests on decades of probabilistic modeling, including hidden Markov models and Boltzmann machines, but its practical form dates from a sequence of architectural advances beginning in 2014. Scale then became the operative variable: models trained on internet-scale corpora with hundreds of billions of parameters displayed capabilities that smaller models did not, which shifted the research question from architecture design toward data curation, training efficiency, and control.

Model Families

Four families account for most deployed systems. Variational autoencoders learn a compressed latent representation and a decoder that maps latent samples back to data. Generative adversarial networks, introduced by Ian Goodfellow and colleagues in 2014, pit a generator against a discriminator in a minimax game, and they dominated image synthesis for several years. Diffusion models, formalized in work on denoising diffusion probabilistic models, learn to reverse a gradual noising process and now underpin most high-quality image, video, and audio generation. Autoregressive transformers predict the next token in a sequence given all previous tokens, and they are the basis of large language models. Hybrid designs are common, such as latent diffusion, which runs the diffusion process inside an autoencoder's latent space to cut computation.

Training, Alignment, and Prompting

A large generative model is typically built in stages. Pretraining on a broad unlabeled corpus with a self-supervised objective produces a base model. Supervised fine-tuning on curated demonstrations then narrows behavior to a task format, and preference-based methods such as reinforcement learning from human feedback or direct preference optimization adjust outputs toward human judgments of helpfulness and safety. Once deployed, behavior is steered at inference time through prompt engineering, the practice of composing instructions, examples, and formatting constraints that condition the model's output distribution. Retrieval-augmented generation extends this by inserting documents fetched from an external index into the context, which grounds answers in verifiable sources and reduces fabricated content.

Evaluation, Provenance, and Risk

Assessing generative systems is harder than assessing classifiers, because there is no single correct output. Practice combines automatic metrics such as perplexity and Fréchet inception distance, benchmark suites, and human preference comparisons, none of which fully captures factual accuracy or downstream harm. Governance work has grown accordingly. The NIST Generative AI Profile, released in 2024 as a companion to the AI Risk Management Framework, enumerates risks specific to these systems, including confabulation, information integrity, data privacy, harmful bias, and misuse for chemical, biological, or cyber capability. Content provenance efforts, including cryptographic signing of media metadata and statistical watermarking of generated text and images, aim to make the origin of an artifact checkable. Annual measurement of capability, cost, and adoption is tracked in the Stanford HAI AI Index report.

Applications

Generative AI has applications in a wide range of fields, including:

• Software engineering, including code completion, test generation, and migration
• Drug discovery and materials design through generative molecular models
• Engineering design, where generative methods propose structural geometries under constraints
• Media production for image, video, music, and voice synthesis
• Synthetic data generation for training where real data is scarce or restricted
• Customer support, translation, and document summarization