Is artificial intelligence a step on a long road or a leap forward?
I think that depends a lot on its practical applicability. In many areas, AI is replacing Google searches, representing significant progress for many people in their everyday lives. Researchers also benefit from AI in work steps that are now becoming more efficient. AI is not only used for complex scientific tasks – such as improving climate models or predicting the three-dimensional structure of proteins – but also to facilitate everyday tasks such as searching. AI models are precisely tailored for this purpose. They help us find and implement ideas more quickly and perform repetitive tasks efficiently. Certain fields of research benefit significantly from this, especially those that involve processing large amounts of data. However, when we consider the long-term development of scientific and technical systems, it is more of a long chain of steps: increasing computing power, theoretical work, internet-based data processing, and digitization, which facilitates data collection. Taken together, these are many small but significant steps toward more advanced AI.
Do intelligent computer systems fundamentally change research?
In climate research and other disciplines, AI is not currently doing anything fundamentally new. Many tasks that previously took much longer, such as data analysis or certain modeling steps, can now be completed more efficiently. Purely AI-driven weather models are currently on the rise and are many times faster than physics-based models. However, I have my doubts as to whether such AI models are currently capable of doing what we are trying to achieve in many areas of science: namely, learning the laws of nature and understanding causal relationships – for example, in climate modeling. This task is based on a vast body of experience and knowledge about the behavior of liquids and gases in the oceans and the atmosphere. On this basis, we simulate climate change and hopefully develop measures to address it. AI can be very helpful in many areas, for example in weather models, but this success principle is currently reaching its limits. This is because weather prediction often involves forecasting conditions that have already occurred, at least in a similar form. Climate change, however, is leading us into completely new situations.
Why can’t AI do this?
Modern AI systems focus on learning correlations from extensive texts and other data sets. They are becoming increasingly computationally intensive and are able to identify correlations, but they do not attempt to find laws or their simple mathematical descriptions. It is a miracle of nature that it can be described and explained through the language of mathematics. However, AI cannot explain why it makes a particular prediction. While there are approaches to the retrospective explainability of trained AI models, these lack fundamental logical and rule-based concepts. An example: modern AI cannot truly learn the rules of chess simply by observing chess games. Although it may master the game very well by processing millions of examples, it still violates rules because it is not programmed to learn them. In contrast, children understand the concept of chess and that they must follow the rules. General AI systems struggle to infer laws because such laws are not embedded in their neural architectures.
Another rather curious example of this is the faulty generation of images of hands with six fingers. When an AI is tasked with generating images of different people, it often produces beautiful images that may, however, be incorrect. The reason for this is that AI does not understand the concept of anatomy and thus the correct number of fingers. The problem is often solved by human intervention, but this is mere error correction – without the computer system having understood concepts and rules.
Why not combine computer systems that learn from large amounts of data with those that are given rules?
One approach to this is causality, but the topic is challenging and requires in-depth theoretical work. Existing AI models were created to process large amounts of unstructured information. To do so, they often represent information in text modules or embedding spaces and use them to calculate probabilities. They lack the essential logic necessary for understanding causality. My area of expertise is causal inference, which is the ability to learn causal relationships from data. However, the methods used so far are not yet really capable of handling large amounts of unstructured data.
Why do we need causal inference?
It allows us to identify and quantify causal relationships from observational data using algorithms and assumptions about the underlying system. You could say that current AI models answer questions such as "What patterns occur frequently?" or "What is the next most likely value or word?" The questions we focus on are of a completely different nature. For example, we ask: Why does a certain weather phenomenon exist, and how is it influenced? Why does an extreme event occur? And we want to know how likely extreme events are to occur in 100 years under the conditions of climate change. These questions are causal in nature because they don't simply ask how often something occurs, but what happens when certain conditions are present or changed. These types of questions arise in the context of system interventions, such as in medicine: "What happens when I take a certain medication?" That is a causal question. We want to know whether aspirin works, not just what its use correlates with.
In other words, AI has to learn to ask the right questions?
The ability to formulate such questions forms the basis for developing a model that is actually capable of answering them. In many areas of research, these are physical models. In addition, one of the significant research achievements of the last 40 years has been the introduction and mathematical description of causal networks that are capable of answering “what if” questions, as well as the development of corresponding algorithms. In order to learn causal relationships from data, it is not always necessary to deduce physical laws of nature. It could also be causal networks based purely on qualitative and probability, for example in medicine, biology, sociology, and also in climate research. An important principle in detecting causal networks is, for example, the assumption that all statistically measured correlations in the data arise from causal relationships and, conversely, that all causal relationships lead to statistical dependencies. Depending on the method, more specific assumptions are also helpful, such as whether relationships are linear or non-linear and what the data distribution looks like – for example, whether it is normally distributed or not. These specialized assumptions are crucial for working with the data and for extracting causes and effects.
Where could causal AI models be used in research?
Basically, anywhere where it is about causal questions rather than pure pattern recognition and frequency analysis. In genetic research, for example, there are experiments in which certain genes are activated to investigate their effects on organisms. Such experiments are extremely expensive. At the same time, we have access to extensive data from previous experiments and observations showing which genes are associated with certain characteristics in nature. Causal models can be used to predict which gene combinations are likely to have a causal effect. This allows us to identify the most promising experiments.
By contrast, there is no possibility of direct intervention in the climate sector. Intervention is possible in cost-intensive physical models of the climate system, but these cannot represent all processes down to the last detail. Causal inference can be applied in many ways here. On the one hand, climate models can be evaluated and improved using causal methods. On the other hand, observational data can be analyzed directly, for example, to understand how the Gulf Stream influences the development of sea ice in the Arctic and what effects this, in turn, has on our weather. We have extensive measurement data from past decades, including information on air and water temperature, pressure, and sea ice extent. By using causal methods, we can use this data to assess which scientific hypotheses are more plausible and gather more evidence for one assumption or another. These methods make it possible to go beyond simple correlations and rule out random or spurious connections. Although we cannot always confirm causal relationships with absolute certainty since the assumptions underlying the methods are not always fully met, we are able to exclude many non-causal explanations. This is the contribution these methods can make to research. To do this, however, causal inference must first learn to deal with large amounts of unstructured data; only then could one speak of “causal AI”.
Jakob Runge has been Professor of Artificial Intelligence in the Sciences at the University of Potsdam since 2025.
Jakob Runge already works closely with scientists from many disciplines. He intends to continuously expand this interdisciplinary position in the coming years – including through an AI Competence Center that aims to integrate these methods broadly into research and teaching at the University of Potsdam.
Causal inference is a field of research that develops theories and methods for learning cause-and-effect relationships from data. Instead of merely describing correlations, it attempts to clarify whether a factor actually causes something or merely occurs at the same time by coincidence.
This text was published in the university magazine Portal - Zwei 2025 „Demokratie“. (in German)
Here You can find all articles in English at a glance: https://www.uni-potsdam.de/en/explore-the-up/up-to-date/university-magazine/portal-two-2025-democracy