Foundations
Generative AIGenAI
Also known as: generative artificial intelligence
AI that creates new content, text, images, audio, video or code, rather than only analysing or classifying existing data.
Generative AI is the kind of AI that makes things. Ask it to write an email, draw a picture, compose a jingle or draft some computer code, and it produces something brand new that did not exist before. This is the technology behind tools like ChatGPT and image generators. It is different from older AI that mainly sorted or recognised things; this one creates.
Generative AI (GenAI) refers to models that generate new content rather than simply labelling or predicting a category. Give it a prompt and it returns text, an image, audio, video or code. It is the branch of artificial intelligence that broke into everyday life from late 2022 onward.
Different content, different models
- Text: large language models such as GPT, Claude and Gemini.
- Images: diffusion models such as those behind Midjourney and Stable Diffusion.
- Audio and video: speech, music and video-generation models.
Many newer systems are multimodal, handling several of these at once.
How it "creates"
It does not copy from a database. It learns the statistical patterns of its training data and then samples new combinations that fit those patterns, which is why every output can be a little different.
Generative AI denotes models that learn a data distribution and sample from it to synthesise novel outputs, in contrast to purely discriminative models that estimate a decision boundary. The current landscape is anchored by two families: autoregressive transformers for sequential data (the basis of LLMs, modelling p(x) as a product of next-token conditionals), and diffusion models for continuous data such as images (learning to reverse a gradual noising process; Ho et al. 2020). Earlier generative approaches, GANs and VAEs, remain relevant but have been largely displaced for frontier text and image work.
The practical significance is that a single pretrained generative model can be steered at inference through prompting, few-shot conditioning and tool use, rather than requiring a bespoke model per task. Key open concerns are factual grounding (mitigated by retrieval), controllability, provenance/watermarking, and the hallucination failure mode inherent to likelihood-based sampling.