Intelligent machines are mastering more and more challenges that we once thought only humans capable of: Algorithms recognize images, understand language, translate texts and even in complex games such as chess or Go, they no longer give their human counterparts a chance.

But how does the rapid triumph of artificial intelligence affect the research of economists, who usually deal more with models for the labor market than with image recognition?

What at first glance has little to do with the methods of economics turns out to be a modern tool kit that has begun to permeate economic research.

To understand the new tools artificial intelligence offers economists, we must first understand how smart algorithms achieve their remarkable successes.

Suppose we wanted to write a program that would recognize writing and turn it into digital text: Gone are the days of trying to hand-program computers to recognize each letter and number.

Instead, we now let the artificial intelligence search for solutions itself: Instead of detailed rules, we now feed the algorithm with data in which text is already assigned to the documents.

So-called machine learning algorithms then automatically extract the connections between scanned pixels and their meaning from this data and transfer them extremely successfully to new documents.

Many applications of modern artificial intelligence are essentially based on statistical models and methods that extract complex patterns from existing data.

Except that these models are no longer assembled by humans, but selected automatically by the machine.

Causality versus correlation

At first glance, artificial intelligence could provide economists with new tools for our empirical analyses.

After all, as economists, we spend a considerable part of our work using our economic models and statistical methods to analyze complex data in order to use it to describe economic relationships.

But the analogy between artificial intelligence and economic modeling quickly reaches its limits.

Because the statistical methods of machine learning focus entirely on making predictions for missing data, such as recognizing the text contained in new documents or deciding whether an incoming email belongs in the spam folder or not.

So you solve prediction questions.

Economists, on the other hand, usually ask themselves questions about understanding the connections in the same data.

For example, given the right data, a machine learning model could reliably map wages based on information such as age, work experience, and education.

As economists, however, we are also interested in why certain wages are particularly high, whether wages are determined in a gender-equitable manner, or how labor market reform would affect wage inequality.

However, machine learning models often make it difficult for us to answer such questions of understanding because of their enormous complexity and lack of economic logic.