At Cafe Hayek, Don Boudreaux points us to a wise quote from Milton Friedman. Below is a comment I left on that post, expanded:
In the highly stylized world of models, where information is perfect, markets are costless, where all preferences are known, where government is costless, and things never change, it is trivially easy to come up with exceptions to free trade and free enterprise. Shift a curve here, refuse to count costs there, and boom! a theoretical reason why tariffs or export subsidies can be beneficial.
However, when those stylized assumptions are relaxed, in other words in a more realistic world where information is imperfect, markets have transaction costs, where preferences are revealed, where governments have administration and operation costs, and where things change, these theoretical reasons disappear like a shadow in the sun. Conversely, the case for unilateral free trade becomes stronger, since it is not dependent upon those assumptions the way the other theoretical cases are; free trade is formulated under those assumptions, yes, but it is robust to movements away. Things like optimal tariffs are formulated under those assumptions but are not robust to movements away from those assumptions.
The true test of any theory is not how well it holds up in perfect conditions, or how well does it perform in the circumstances in which it was conceived, but how robust it is to movements away from those idealized conditions. Economists from Adam Smith to Harold Demsetz and beyond have warned us against these nirvana fallacies. True knowledge is gained when we stress-test our models and see how robust they are. Testing this robustness gave us such fields as Public Choice, Law & Economics, Political Economy, Money and Banking, and the like.
Economic models serve a purpose: they are ways of thinking, methods of analyzing phenomena. However, they are not descriptive of reality. They were never meant to be. When basing policy off of those models, the policy-proponents are making a grave mistake: they are moving their models away from the abstract and into the descriptive. In other words, they are taking their models too literally. This literal interpretation of models can be extremely dangerous.