16.06.2025

What is Artificial Intelligence?

What is AI? – Dr Colin Paterson (SAINTS Training Co-Lead) explains…

For years, we have known the value of creating and using models. We create models for all kinds of things, predicting the weather, modelling airflows, patterns of movement in crowds, and in traffic.

In all cases, we think about how the real world works and write mathematical equations and rules which allow us to map a set of inputs to a set of outputs. For example, we might measure temperature and atmospheric pressure over a few days and use this to predict what the weather might be like tomorrow.

This is fine when the relationships are well understood and the mathematics or rules we use in our models can be well defined. But it usually takes a great deal of expertise or domain knowledge to create models which are useful. Unfortunately, for some problems, the real world is too complex to understand at the level of detail needed to create a model which is useful. Indeed, even when creating a model is possible, the cost of creating it may be prohibitive.

So, wouldn’t it be great if we could just get the computer to work out what the model is for us?

Well, to some extent, this is what artificial intelligence is doing for us.

Rather than working out the mathematics and rules of a model, we can just provide data in the form of examples to show what the inputs might look like, and what the output should be in response. The computer then slowly changes the parameters of a mathematical model until the outputs look like what we would expect for the data we provided.

And this approach has been tremendously successful. We can predict house prices, the occurrence of cancer, or what the next word in a sentence might be. All we need is enough data and a model flexible enough to represent the problem we are interested in.

AI appropriately deployed can solve problems which are more complex than traditional methods and do so in a fraction of the time that a human would take. In a world where resources are limited and time is critical, then AI might well allow for solutions which would otherwise be impossible.

So what’s the problem?

Well, our old approach of model construction required us to engage with the problem, to deeply understand the nature of the models as well as the limitations and assumptions which underpinned the results produced by the model. By short-circuiting this process, we lose this deep understanding.

And when the models and the problems they are looking to solve are complex, it’s hard to know if the solution presented is right or just plausible. Indeed, if I show you 20 correct solutions in a row, you are going to start to believe that the model is always correct, but maybe all those 20 problems were easy and just like the data on which the model was trained. Problem 21 might be unusual and poorly represented by the training data.

Maybe this is OK when we are asking for fashion advice from a model, but less so when the output for the model is part of a larger safety-critical system.

To make matters worse, the world is dynamic and changes constantly unlike our training data, which captured a historic state of the world. Without a concerted and considered approach to mitigating the effects of such change, can we be sure that the systems we build continue to be safe after deployment?

The cost of AI may, therefore not be in pounds sterling, but in lost knowledge and increased risk leading to a loss of safety guarantees. Is that a cost we are willing to pay?