What is Generative AI?
Generative AI is a category of artificial intelligence that creates new content. That content might be text, images, audio, video, code or 3D models. The word "generative" distinguishes these systems from the older, more common class of AI that simply classifies or predicts.
To understand why this distinction matters, you need to understand two fundamentally different ways AI systems can approach the world.
Discriminative vs Generative Models
Discriminative models learn to draw a boundary between categories. Given an input, they answer: which class does this belong to? A spam filter is discriminative: it reads an email and outputs "spam" or "not spam." A fraud detection model is discriminative: it takes a transaction and decides whether it is fraudulent. These systems are excellent at classification and prediction. They do not create anything.
Generative models learn the underlying structure of the data itself. Instead of learning a boundary between classes, they learn what the data looks like: the distribution of pixels in images of cats, the statistical structure of English sentences, the patterns in musical compositions. Once they have learned that structure, they can sample from it: producing new outputs that look like they could have come from the same distribution as the training data.
The technical way to say this: a discriminative model learns P(class | data), the probability of a class given the data. A generative model learns P(data), the probability distribution over the data itself.
What "Generating" Actually Means
When a generative AI model produces an image or writes a sentence, it is not retrieving something from a database. It is sampling from a learned probability distribution.
Consider text generation. A language model has learned the statistical structure of language from billions of examples. When you give it the prompt "The capital of France is", the model assigns a very high probability to the next token being "Paris" and a very low probability to tokens like "elephant" or "seventeen". Generation is the process of repeatedly sampling the next token from this distribution until a complete output is produced.
This probability framing has important implications:
- The same prompt can produce different outputs on different runs, because the model is sampling
- Temperature controls how "peaked" or "flat" the distribution is: a low temperature makes the model more deterministic, always picking the highest-probability token; a high temperature introduces more randomness and creativity
- The model has no ground truth to look up: it is interpolating from patterns learned during training
The Generative AI Use Case Landscape
Generative AI is not one thing. It manifests across several distinct modalities:
Text generation Large language models like GPT-4, Claude and Gemini generate coherent, contextually appropriate text across an enormous range of tasks: answering questions, writing code, summarising documents, translating between languages, composing emails, explaining complex topics.
Image generation Systems like Stable Diffusion, DALL-E and Midjourney generate photorealistic images, illustrations and artwork from text prompts. These systems can create images of scenes, objects and people that have never existed.
Code generation Models like GitHub Copilot, Claude and GPT-4 generate functional code from natural language descriptions or by completing partial code. This is transforming software development: studies suggest developers using AI code generation are measurably more productive.
Audio generation Systems can generate music in specific styles, create realistic-sounding speech from text (text-to-speech) and even clone a specific person's voice from a short audio sample.
Video generation The newest frontier. Systems like Sora (OpenAI) and Runway can generate short video clips from text prompts. The quality is advancing rapidly.
3D and molecular generation Generative models are being used in scientific domains to generate novel 3D molecular structures for drug discovery (an area where AlphaFold made landmark progress), material designs and 3D assets for games and virtual environments.
Why Now?
Generative AI has existed in research form for decades. What changed?
Three things converged simultaneously: vastly more training data (the digitisation of human knowledge at internet scale), dramatically more compute (GPU hardware and cloud infrastructure) and architectural breakthroughs (the Transformer, diffusion models). The combination unlocked a qualitative jump in capability.
The result is a technology that, for the first time, can produce outputs that are genuinely indistinguishable from human-created content across a widening range of domains.
Quiz: What is the difference between a discriminative model and a generative model? Explain the "probability distribution" framing of text generation in your own words.