When AI Becomes Fieldwork: Human–AI Interaction for Coding Moral-Elite Biographies

LLM and AI constitute today a booming field of research also for sociologists and their impact and usefulness in speeding some labor-intensive aspects of our work should not be understated but not overstated either. In this presentation, I discuss the process of prompt engineering that my project team is currently developing in order to exploit LLMs for coding a large corpus of textual data into a prosopographical dataset of career events (Lunding et al. 2020) Our focus is on “moral elites,” conceptualized as individuals who possess resources and positions enabling them to influence central societal norms. Our sample includes actors identified in secondary sources as active in public debates on welfare-state arrangements across diverse subsectors, from civil society and expertise to politics and social and traditional media. This mixed reputational and quasi-decisional sampling (Hoffmann-Lange 2018) quickly pushed us to seek assistance from an AI research companion. However, the heterogeneity of these elites and the non-institutional, often ambiguous nature of their career trajectories requires careful reflection when using LLMs. Unlike political elites, moral elites experience career events that can be blurred, informal, and embedded in narrative flows, raising the risk of arbitrariness in defining and coding them. Accused of being “stochastic parrots”, AI chatbots operate through probabilistic patterns, and their outputs—particularly in proprietary settings—can be highly unpredictable. For this reason, they cannot be deployed unmonitored but require sustained and systematic Human–AI interaction Following use of generative AI in qualitative coding (Ramanathan et al. 2025), we tried to build a structure of Human-AI interaction that envision human active intervention before, during and after the coding process left to AI, in an iterative process that resembles that of grounded theory towards the interaction between theory and data, but treats AI as the fieldwork, from which learning and successive adjustments are needed. The construction of a detailed codebook is of paramount importance and starts necessarily with the contextual knowledge of human coders pursuing an inductive coding strategy based on a selection of biographies and oriented towards outcomes that could be compared longitudinally. Categories should present a detailed description, including a scheme of “trigger words” that could help AI in identifying them. A preliminary instruction for the model is to segment each biography into individual sentences and then assess, sentence by sentence, whether a career event is present. Once this detailed codebook is ready, AI is tested on some biographies, starting a process of learning for both the AI and the researcher supervising it, with a particular attention to categories holding a specific meaning for different historical periods. Memorization features of some AI allow to track down frequent errors, which should indeed be incorporated in the codebook to refine prompts. Treating interaction with AI as temporary fieldwork establishes a learning process that ends up in an adjusted prompt that, once memorization is deactivated, could be applied to the whole sample, relegating human intervention only to the ex-post evaluation of results, thus considerably speeding up the data collection process.