What is it about?
Understanding how one thing causes another in a complex system, like how freedom of speech affects fair elections in a democracy, is essential for science and policy, but hard to do when the underlying relationships aren't known in advance and change over time. This paper introduces DCNAR, a two-step AI method that first learns which variables actually influence each other from real-world data, then tracks how the strength of those influences evolves. We test it on democratic institutions across 139 countries and show that it produces more stable and interpretable answers than existing approaches, even when data is limited. The goal isn't just better prediction; it's giving researchers a reliable tool for asking "what if?" questions about systems that change.
Featured Image
Photo by Logan Voss on Unsplash
Why is it important?
Most existing methods for studying how causes and effects change over time require researchers to assume the relationships in advance. In fields like political science, ecology, and social science, that assumption is often the very thing under investigation. This work is unique in treating the causal structure itself as something to be learned from data and tested, rather than fixed at the start. That shift matters because it opens dynamic causal analysis to areas where it hasn't been reliable, from tracking democratic decline to studying institutional change, giving researchers a way to explore how complex systems evolve without pretending to know the answers beforehand.
Perspectives
Much of the excitement around AI right now is about prediction: better forecasts, faster answers, more scale. That is not what interests me. What I care about is whether AI can help us understand systems we don't yet understand, particularly the human ones. DCNAR is a small step in that direction: a method designed less to predict what will happen than to help researchers ask better "what if?" questions about the systems they study. I hope it contributes to a broader shift in how we think about AI in science: not as an oracle, but as a careful instrument for reasoning under uncertainty.
Valentina Kuskova
University of Notre Dame
Read the Original
This page is a summary of: From Causal Discovery to Dynamic Causal Inference in Neural Time Series, August 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3770855.3818956.
You can read the full text:
Resources
Contributors
The following have contributed to this page
