Research

AI-generated podcast with NotebookLM summarising some key points of the paper: Link

Notebook to try own AI-led interviews within minutes and full platform code available at: https://github.com/friedrichgeiecke/interviews

The advent of large language models (LLMs) provides an opportunity to conduct qualitative interviews at a large scale, with thousands of respondents, creating a bridge between qualitative and quantitative methods. In this paper, we develop a simple, versatile approach for researchers to run AI- led qualitative interviews, including voice interviews. We assess its robustness by drawing comparisons to human experts and using several respondents-based quality metrics. The versatility of the approach is illustrated through four broad classes of applications: eliciting key factors in decision making, political views, subjective mental states, and mental models of the effects of public policies. High performance ratings are obtained in all of these domains.

Challenging dynamic decisions are ubiquitous in economic policy making. At the same time, powerful methods that allow solving dynamic problems of previously prohibitive complexity have been driving much of the progress in artificial intelligence. Yet, they have so far been used mainly in games, robotics, theoretical models, or most recently reasoning in language models. We show that the same methods from reinforcement learning can be utilized to learn optimal policies in novel economic simulations built with existing real-world data and causal inference. Such policies may be able to support economic decisions in extensive dynamic environments with budget and capacity constraints. We allow for restricted classes of policy functions and prove that their regret decays at rate n−1/2, the same as in the static case. Illustrating our approach with the complex dynamic allocation of job training and reskilling programs in labor markets, we find that by exploiting the problem's dynamic structure, we achieve significantly higher welfare compared to static approaches.

How do political careers relate to family networks? We address this question using a novel dataset assembled from various sources and combining information on the political careers of British MPs with information about the family trees of the British aristocracy. Our data span two centuries (1832-2022) and document the evolution of backgrounds of British MPs over time, their social origin (in particular whether aristocratic or not), education (university and school attended), profession and political experience. Using information on family trees we can then calculate the centrality of those MPs who are part of the aristocratic network and assess the role of network centrality in election and career progression.

We use methods from natural language processing and machine learning to characterize the innovative content of patents. We elaborate the concept of widening innovations, which are patents whose textual content is different from existing patents at the time when they are filed and similar to patents subsequently filed. The term “widening'' reflects the idea that these patents extend knowledge in a novel way and spur innovation. The computations of patents' similarities rely on numerical representations of the textual content of patents. We show that widening patents have higher citations and the firms owning them make higher profits and grow faster relative to other firms, although causality between filing such a patent and firms' performance cannot be established. Unsurprisingly, many widening patents stem from the IT revolution, a field near-non-existent in the 1980s which became predominant in the innovation landscape.