It’s behind the scenes in our favourite apps, making breakthroughs in the lab, and constantly in the news. Artificial Intelligence (AI) is all around us, but how does it work? Let’s shed some light on the world of AI.

‘Artificial Intelligence’ is an umbrella term that covers a wide range of systems, all working in slightly different ways. However, many of the most widely known AI systems use mathematics to find patterns in data. These patterns are then used to make predictions. The most common forms of these AI systems use something called machine learning, where algorithms analyse data, arranging the patterns and features into models.

You can think of models like a map. Let’s consider how a machine learning model might use an image of a koala. This image contains millions of pixels. When we put this digital file through a machine learning model, these data points are processed via many layers of multiplication and addition until patterns start to emerge for different features. You can think of these features like islands. The more images we add, the more comprehensive the map becomes. For example, on ‘Koala Island,’ the west side represents koalas with small ears, and the east side represents koalas with big ears.

What’s really mind-blowing is that ear size is only one feature. To represent all the different shapes, colours, moods, and compositions of koala images, we not only need to work in 3D but also imagine thousands of dimensions. For now, let’s keep it to just 2! Now, if we take this map, which has been trained with millions of koala images, and ask what an image might look like… an AI system can generate a completely new image related to that location.

This is called Generative AI. Amazingly, this mapping process works for any data—whether it’s text, images, sound, or anything that can be described with numbers. When we train models with different types of data together, it’s a bit like combining two maps. For example, here we have a text-to-image model trained using images and their corresponding text labels.

These models can take on complex tasks, answer questions, write poems and music, and even generate videos from scratch. But fundamentally, what we’re doing is giving computer systems a way of mapping information and making connections between patterns. It’s maths, not magic. While these outputs are often convincing, it’s important to remember that they are only predictions based on training data.

For example, if we go back to our text-to-image map, and imagine that our text map was trained using American examples, instead of Australian, this prompt might give us a much different result. Sometimes, this can lead to something called bias, where unfair or unbalanced outputs are generated that amplify inaccuracies or gaps in the data.

Take planning a playlist, for example. If an AI was trained using only your listening history, it wouldn’t have the right data to generate a playlist that appeals to everyone. It’s important to remember that human and artificial intelligence are different. For instance, humans instinctively understand context and apply common sense, while AI systems approach things differently.

That’s one of the reasons it’s important to understand how AI systems work—to ask questions and decide when and how we want to use them. In the right hands, AI systems can be incredibly powerful tools, helping manage huge data sets, see patterns that humans can’t detect, and automate complicated processes. But it’s up to all of us to ensure AI is directed toward the brightest future possible.

To find out more, visit csiro.au/ai.