In the philosophy of science there is a debate about whether scientific theories tell us what the world is really like, or whether scientific theories are nothing more than ‘tools’ or ‘instruments’ – useful for making predictions, but not for telling us what the world is really like. This debate is called the realism-debate. Most working scientists are realists: they believe that scientific theories tell us what the world is really like. They argue that realism is the only philosophy of science that can explain the success (in terms of the accuracy of predictions) of science. For example, they argue that “the theory of atoms allows us to predict that a gas expands when it is heated; wouldn’t that be a mystery if atoms did’t exist?”
Convergent realism – do better predictions lead us truth?
Most realists would admit that we can never be 100% certain that our scientific theories tell us what the world is really like. “But”, they argue “science gets better and better at making predictions, and that shows us that we’re on the right track – we’re getting closer to the truth.” The idea that our realists defend is that of convergent realism. The assumption of convergent realists is that as science progresses, it gives us a better picture of what the world is really like. Often without being aware of it, convergent realists assume that better prediction goes hand in hand with a more truth-like description (because that’s what ‘being on the right track’ means).
We’ll see that the realist’s assumption that better prediction implies a more truth-like description is problematic. Consider, as an analogy, a map of some city-centre, on which all the streets and buildings are represented by lines on a flat surface. We will call the city map a ‘model’ of the city. Imagine that there is a second model of the city, and that the walls in this second model are made out of real stone (as perhaps in a scale model); they are not flat lines on a piece of paper as the walls in the first model are. Suppose also that the second model is very incomplete and that it has only several streets and buildings in it. Now we ask: which model is better?

Scale model of ancient Rome: even if the description contains all roads it doesn’t tell us what happens when we bump into a wall.
Better predictions lead to… better predictions
We are going to test the models. Suppose that we are actually in the city-centre and start walking in a certain direction. Suppose that both models tell us that there’s a wall ahead. I don’t know about you, but I’d like to be able to predict what actually happens if I cross something that looks like a line on the map! Okay, perhaps the first model has the advantage that you can use it anywhere in the city-centre (because it’s complete), but the second model helps you predict what actually happens if you bump into a wall (because walls in the model are vertical stoney structures). We see that the question of which is the better model is not easy to answer. If we simply assume that the best model is the one that yields the best predictions, then the scientist’s claim that better predictions show that we are closer to a true picture (the perfect model) becomes a tautology: better predictions show that we have a better model, which, in turn, means nothing else than making better predictions. So better predictions show us that we make better predictions – we don’t have to be convergent realists to know that.
What goes for city-maps also goes for scientific theories: some of them are great for making predictions, but that does not mean that they tell us ‘what the world is really like’. This shows that if scientists say that the better their theories predict, the more accurately these theories represent reality, they’d better scratch their heads some more.

Street map of ancient Rome: not only is the map incomplete, it also doesn’t tell us what happens when we bump into a wall.