Large Language Models (LLMs) have rapidly become one of the most relevant scientific topics, both as theories of language and cognition, but also as tools for research. This trend has accelerated even more since applications like ChatGPT are available to the public. This course will first give a basic introduction to LLMs, outlining the core components of their architecture and how they are trained. For this we will take a bird’s eye perspective, focussing on the key concepts rather than mathematical details: The training data, the training objective (predicting words in contexts), basics of (deep) neural network models, simple language models, the central components of the transformer architecture that underlies common LLMs, and the difference between unidirectional and bidirectional models. This part of the course will end with an overview of how users can interact with LLMs beyond chatbot interfaces (for example, via API calls), and how we can access not only their outputs, but also some of their internal representations and weights. 

The second part of the course will focus on use cases of LLMs in research, in their role as “research assistant systems” that allow us to automatize and “outsource” labor-intensive tasks. This will include using LLMs 

• as “pseudo-participants” to pilot or simulate behavioral studies
• to automatically generate and simulate new datasets (or extrapolate and scale up existing datasets)
• to automatically annotate data and automatize qualitative analyses
• to analyze open text responses, thus taking a quantitative rather than a qualitative approach to text analysis
• for systematic literature reviews 

Prerequisites: The course is targeted towards a general audience with no specific pre-existing expertise in the topic. General coding skills are useful, but knowledge of a specific programming language is no requirement for participation.

Many questions in the social sciences, public health, economics, and business, such as whether a policy works, whether an intervention reduces risk, or which patients benefit most, are fundamentally causal. Yet modern datasets are increasingly high-dimensional, messy, and rich in potential confounders, making purely parametric approaches brittle: small modeling mistakes can translate into large bias in estimated treatment effects. At the same time, many machine-learning methods excel at prediction rather than causal estimation, so naively plugging ML models into causal workflows can yield strong accuracy metrics while still producing misleading effect estimates.  This course presents an applied, end-to-end workflow for causal inference with AI/ML methods, emphasizing reliable treatment-effect estimation and valid uncertainty quantification. We begin by formalizing causal questions using the potential outcomes framework and causal graphs, emphasizing identification: what can (and cannot) be learned from observational data under explicit assumptions. We then move to estimation in complex settings and introduce debiased (double) machine learning (DML), a principled framework that leverages flexible ML for the components needed to estimate treatment effects while supporting valid inference. Building on this foundation, we cover advanced topics that practitioners routinely face: heterogeneous treatment effects and how to summarize and communicate effect heterogeneity responsibly; robustness and sensitivity analysis to assess how conclusions change under plausible violations of assumptions. If time permits, we conclude with a brief tour of current research directions at the intersection of causal inference and AI. The course is intended for applied researchers and data scientists interested in estimating causal effects.

Background reading: causalml-book.org