I’m warning you, I might get more philosophical than usual! I’ve been thinking recently about how artificial intelligence (AI) has prompted some questions that are of immense philosophical interest to me. However, I know that such questions often seem obscure to the outsider. So beware – and hear me out!
To be fair, I was nudged by my editor. She told me about the very exciting Chemistry World article on how AI’s predictive accuracy of catalysts is currently refined by training it on experimental measurements and data. And, it just so happened that a couple of weeks ago I was talking at a workshop about exactly that: namely, how are AI models trained in chemistry and what does this tell us about the truth of scientific theories?
For all this to make sense, I need to offer some context first. In philosophy of science, it is standard to invoke the success of scientific theories in support of their truth. The argument, called the ‘No Miracles Argument’, goes like this. Our best current theories are very good at predicting facts, explaining phenomena, and helping us understand the world around us. They are used to make planes that stay up in the sky, cure diseases, and so on. All this success would be miraculous if these theories weren’t true. That is, if it were the case that our scientific theories do not track (at least) something accurate about the world, what it contains, and how it is structured, then it would be quite inexplicable how they manage to be so successful. This is one of the strongest arguments that scientific realists invoke to argue in favour of the truth of our best current theories.
The power of predictions
This argument goes hand in hand with another idea, called predictivism. This idea is about the evidential weight that prediction holds, not just in supporting the truth of a scientific theory, but also when choosing among competing theories. Specifically, philosopher Karl Popper noticed that when scientists have to choose between theories that are equally consistent with empirical evidence, they tend to pick the one that most successfully makes predictions. But not all types of predictions will do – they have to be truly novel. Here, a very important distinction is made between accommodating known facts and making novel predictions.1
Take, for instance, the periodic table. As I’ve discussed elsewhere, by 1870 there were several ways to classify chemical elements. For some time, there was some indecisiveness as to which classification should be used as pretty much all of them accommodated the known elements relatively well.2 What flipped the scale in favour of Dmitrii Mendeleev’s version was that it managed to do something the others didn’t: predict previously unknown elements.
This is a classic example of how scientists choose based on the capacity for novel predictions. More importantly, it shows how crucial such predictions have been thought to have been when supporting the truth of theories and theoretical schemas.
But how does AI affect all this? AI models are currently developed to predict new stuff. In chemistry in particular, they aim at predicting the properties of known substances – what we would call accommodating known facts. But they are also designed to predict novel substances! The interesting thing about this is that these AI models are not always trained on a theory. In fact, my preliminary research has revealed that many AI models are trained either on data alone (drawn from experiments and measurements), or by inserting very few theoretical constraints with the data. There are also AI models, like the one developed at the University of Bristol called Impression, which are trained using data that is computed through quantum chemical models.3
Truth in theory
If AI succeeds in making novel predictions without having to ‘learn’ any theory, then it seems we have a philosophical problem. Either philosophers were wrong to put so much value on predictive success when developing arguments in favour of the truth of scientific theories. Or – and this is worse – they cannot be so sure about the truth of scientific theories. This is because we can no longer invoke predictive success as safe evidence for a theory’s truth.
One could respond to this that AI is somehow learning the theoretical connections that exist between the data it is being ‘fed’, and which we humans have already discovered through our theories. But how can we know this? Here, the so-called problem of opacity becomes particularly pressing. This problem has to do with the fact that we just don’t know how AI produces the results it does. It acts like a sort of black box whose inner workings are to an important degree mysterious to us.
I have to say though, I’m still getting my head around how all of this really works and there are aspects to it that I still don’t understand. So, I might be worried for nothing! Also, nowadays, changes are happening so fast that even if I’m right now, what I am claiming might become completely wrong in the near future. For example, it seems that there are already efforts to solve the problem of opacity through the development of Explainable AI, namely methods that aim to help us understand how AI produces its predictions. I suppose we have to wait and see. Until then, some philosophical fun never hurt anyone!
