Doing the Research seriesHow Can AI Recognize Causality?
25 June 2025, by Newsroom editorial office

Photo: AI-generated using DALL-E
AI is now part of almost all areas of life. Machine learning methods can help us identify patterns in large amounts of data, but can they also identify and evaluate causal connections? Dr. Philipp Bach, who is working in the team headed by Prof. Dr. Martin Spindler, professor of statistics at the University of Hamburg Business School, is researching this question.
What is the basic question of your research?
Bach: Machine learning methods can help us model highly complex connections, meaning patterns in large amounts of data. Traditional approaches focus on prediction tasks, for example, for buying decisions or diagnosing disease. But we are also looking at questions of causality: Can we use AI models to recognize and depict cause and effect? And then correctly evaluate the causal effects?

Why is that important?
When we make decisions, we weigh the advantages and disadvantages of my different options. For example, we consider questions such as: which sport should I do to improve my health? In politics, the effects of raising the minimum wage on unemployment are relevant. To make good decisions, it’s important to understand causal relationships and their fundamental mechanisms. This is where data help, and where there is a lot of data—for example, in the personalization of treatment in medicine—AI can help.
What is the major challenge?
All causal inquires involve a statistical problem: for example, when approving medical treatments, you want to know if it will really, on average, improve health. Because we cannot examine everyone who is sick and work only with sample data, we can only guess at potential efficacy. This leaves some statistical uncertainty and we are asking ourselves: how precise is our guess? And how can we quantify uncertainty?
So in addition to data we are also searching for the right algorithms?
Yes, because this starting point affects the very flexible models of machine learning, meaning they adapt to complex patterns in the data. This is a problem for us, because our guesswork needs to fulfill certain quality criteria, for example, avoiding systematic distortions. This would lead to incorrect estimates and then to misleading conclusions.
This is why in our field of research we take approaches that make it possible to use the flexibility of machine learning while also ensuring valid statistical estimates of causal effects. A fundamental method is “double machine learning,” which plays a central role in our research group. This requires additional technical skills in software development and our sound technical equipment in the Hummel-2 cluster at the Regional Computing Center, with its many GPUs, also benefits our project.
Can you name, for example, where these AI methods are being applied?
We often work in the field of econometrics, meaning statistics in the economic sciences. In a current research project, for example, we are looking at the role that text and image data play in online shopping. We assume that textual descriptions and product images influence buying decisions from Amazon, etc. Meaning that thanks to certain features, some products seem more aesthetic, interesting, or of better quality and are thus bought more often. This is a causal connection. When I recognize these effects and now which photo motifs and key words work especially well, I can better assess customers’ desire to buy and optimize pricing.
In our project, we have set up a dataset with comprehensive descriptions of toy cars with different features such as image, description, and relevant data on size and price. Then there is information about how many people have bought items with certain features. We can integrate the text and image data into our causal inquiry using the modern architecture of neural networks. This way we can guess which factors are relevant for buyers and to what extent they influence a willingness to pay. Using large language models, we can better understand the role of certain product features in purchasing decisions.
Doing the Research
There are approximately 6,200 academics conducting research at 8 faculties at the University of Hamburg. Many students also often apply their newly acquired knowledge to research practice while still completing their studies. The Doing the Research series outlines the broad and diverse range of the research landscape, and provides a more detailed introduction of individual projects. Feel free to send any questions and suggestions to the Newsroom editorial office(newsroom"AT"uni-hamburg.de).

