What is Generative AI?

Generative AI is a class of Artificial Intelligence that is able to create new and 'original' digital content. This includes text, images, video, music and computer code. Generative AI works by analysing patterns in large sets of training data and encoding them into predictive mathematical models; a machine learning process commonly referred to as training. Importantly, Generative AI models do not keep a copy of the data they were trained on and then simply reproduce it when requested; rather they generate new content from the patterns they have learned.

Generative AI tools typically provide users with a web-based chatbot interface to input a prompt – an instruction in plain, natural language. This initial prompt, and subsequent further conversational replies, are what instruct Generative AI models to create new content.

Flowchart showing generative AI as a brain, taking in a prompt and producing outputs.

Generative AI is developing and getting better all of the time, and is proving to be a helpful tool for a range of different tasks in daily life, education and in the workplace.

Large Language Models

Generative AI tools use neural networks (a type of computer programme that emulates how a brain connects pieces of information) to identify and 'learn' the patterns and structures within existing data. For example, for text-generation AI tools, they will learn that certain words tend to follow others, or that particular phrases are more common in certain contexts. The more data the AI is trained on, the better it will perform and the more accurate and novel its outputs will be. ChatGPT, for instance, is trained on massive amounts of data gathered from millions of books, online articles, databases, Wikipedia pages, and forum-based discussions.

AI neural networks that specialise in text-generation are known more commonly as Large Language Models (LLMs), and are particularly good at understanding, reproducing and conversing in natural language. When given a prompt, the model uses this knowledge to analyse and process the request before producing a completion, or output, which is then presented back to users.

Prompt and completion in a ChatGPT interface, linked by a model (GPT-3.5, GPT-4).

The quality of the generated output will depend on a number of factors, such as the amount and quality of the training data, the complexity of the prompt, and the model's size. Larger models usually generate better outputs but require more computing power and resources.

Generative Adversarial Networks

The other type of AI model you will likely encounter is the Generative Adversarial Network (GAN). These are used to power image, video and music AI generation platforms. Unlike LLMs, GANs use two neural networks that compete against each other as they train. One network (called the “generator”) uses its training to create new content. The second (called the "discriminator”) tries to work out if that image is real or fake. The first network is rewarded every time it fools the second one. The second one is rewarded every time it is not fooled. Both networks are improved, by repetition, until the first network becomes extremely good at generating outputs that are indistinguishable from the training set.

An illustration of the GAN process

©Wang and Pan et al

Notice that the Generator creates content, not from scratch, but from randomly generated "noise" (similar to TV static). The model uses this random noise as a starting point to help it create something new each time it runs - similar to how an artist would look for inspiration by considering new and novel influences.