Large language models (LLMs) are rapidly transforming scientific work, but their integration into research workflows raises profound epistemic, methodological, and ethical questions. This presentation clarifies the distinction between research on AI (examining model behaviour, biases, evaluation reliability, and failure modes) and research with AI, where models function as instruments embedded in data pipelines for coding, semantic search, corpus exploration, simulation, and large-scale automated analysis. Using empirical results from recent studies , I illustrate how LLMs can both enable and distort scientific work.

The core of this lecture is practical: a step-by-step “AI integration pipeline”. Participants will see concrete examples, including prompt design, sanity checks, error inspection, and hybrid human–AI evaluation strategies.

Throughout the lecture, I argue that AI multiplies responsibility: researchers are accountable for the data they expose to models, for the outputs they accept, and for the downstream effects of hybrid human–AI analysis. The goal is not to replace scientific judgment, but to strengthen it leveraging AI’s analytical power while preserving methodological rigor, transparency, and epistemic integrity.