Modern machine learning methods have transformed causal analysis, offering powerful tools for high-dimensional data and complex structural settings. These methods address key challenges such as confounder adjustment and flexible modeling of outcomes and treatments (Ratkovic, 2023; Cinelli & Pearl, 2022; Wager & Athey, 2018; Chernozhukov et al., 2018). The least developed element in practical statistics, however, is the human component: researchers are left to make key choices without systematic guidance or support (Gelman & Loken, 2014; Hill et al., 2024). Experts often publish software and methodological papers, but then largely step back, leaving users to navigate these tools on their own.
Fully automated AI systems seem a natural next step, but current implementations are themselves beset by biases, inaccuracies, and hallucinations. Instead, I propose scaffolding software for AI-assisted causal inference. The aim is to develop interactive AI systems, building on fine-tuned large language models (LLMs), that guide researchers step-by-step through causal and analytic workflows. Specifically, these interactive AI systems will explain assumptions in plain language, log all modeling decisions to address hidden researcher degrees of freedom, and generate reproducible code. The societal relevance of this work will be demonstrated through collaborations with substantive scholars in three domains: sex trafficking, organ harvesting in authoritarian states, and right-wing misinformation. Theoretically, we will focus on theorizing the interface between humans and AI. These domains and theory will serve as testbeds for AI scaffolding, showing how improved causal workflows can lead to better understanding and real-world interventions.
By the end of the project, we will deliver functional prototypes, reproducible pipelines, and a fully drafted ERC Synergy proposal, positioning Mannheim as a European leader in socially responsible AI. This work will be embedded within Mannheim’s growing AI ecosystem, including a new AI Incubator Fellowship jointly run with the Business School. Together with the Chair’s cutting-edge GPU infrastructure, this provides an ideal foundation for scaling both technical innovation and societal applications.
