What Is Generative AI?

For most of computing history, AI systems were built to recognize, classify, or predict. A spam filter decides whether an email is spam or not. A fraud detector flags suspicious transactions. A recommendation engine predicts what show you might watch next. These are extraordinarily useful — but they all work by looking at something that already exists and making a judgment about it. Generative AI does something categorically different. It creates. It produces new text, images, audio, video, and code that did not exist before. When you ask a text model to explain photosynthesis in three sentences, it writes those sentences. When you describe a dragon perched on a lighthouse at sunset, an image model paints that scene. Nothing was retrieved from a database or copied from a file — it was generated, token by token or pixel by pixel, from the model's learned understanding of the world. This module is about how that works, and — critically — how to guide it. By the end, you will understand what is happening under the hood when you type a prompt, why results can be brilliant or baffling, and how to become a skilled operator of these tools.

Two Eras of AI

Think of earlier AI as answer machines: given a specific question with a bounded answer, they deliver that answer. A chess engine evaluates positions and picks moves. An image classifier looks at a photo and outputs 'cat' or 'dog.' A sentiment analyzer reads a review and outputs 'positive' or 'negative.' In every case, the output space is predefined — there is a fixed set of possible answers. Generative models operate in open output spaces. The output of a text model is not chosen from a list — it is constructed word by word (technically token by token, which we will cover in Lesson 3). The output of an image model is not retrieved from a gallery — it is synthesized pixel by pixel. This is what makes generative AI feel almost magical to use and what makes it genuinely hard to reason about: the space of possible outputs is effectively infinite.

The shift from discriminative to generative AI did not happen overnight. The key breakthrough was scale — specifically, training enormous neural networks on enormous datasets. A model trained on billions of web pages sees so much human writing that it develops a rich internal representation of how language works: grammar, facts, reasoning patterns, tone, analogy, humor, and more. An image model trained on hundreds of millions of image-caption pairs builds a similarly deep understanding of visual concepts. This scale is why modern generative AI behaves so differently from older AI. It is not following handwritten rules. It has learned a compressed model of human knowledge and can generate new outputs that reflect that knowledge in novel combinations.

What Can Generative AI Create?

The range is broader than most people realize when they first encounter these tools. Text models can write essays, summarize documents, translate languages, answer questions, draft code, explain scientific concepts, write fiction, and converse naturally. The same underlying technology — a large language model — handles all of these because it has been trained on text spanning every domain. Image models can generate photorealistic photographs of things that never existed, illustrate concepts in any artistic style, edit existing images, or create entirely new visual compositions from a text description. Audio models can synthesize voices, compose music, and generate sound effects. Code-generation models can write, debug, and explain software across dozens of programming languages. These are not separate technologies bolted together — they share deep architectural similarities. Understanding one helps you understand all of them.