IMPACT OF LARGE LANGUAGE MODELS ON THE COGNITIVE PROCESS

Large Language Models (LLMs) are transforming the way we learn and work, yet their impact on human cognition remains underexplored. This study examines whether, and to what extent, LLMs influence users’ answers and confidence in general knowledge questions, comparing two interaction modes: conversational (e.g., ChatGPT) and streaming (e.g., GitHub Copilot). We designed two controlled experiments: a Pilot Study, where participants answer without LLM assistance, and an Agent Study, where they are assisted by either a conversational or streaming agent. Custom-built agents ensure full control over the suggestions, including the deliberate introduction of incorrect answers, enabling the analysis of user behaviour when faced with misleading information. The thesis originated from an observation by Professor Arnaud Legout, a researcher at INRIA Sophia Antipolis and professor at Université Côte d’Azur, who supervised this work. As a frequent user of GitHub Copilot, he noticed a recurring “push” to accept the tool’s suggestions even when they did not match his intentions, which raised a broader question: why do users feel encouraged to accept AI-generated suggestions — convenience, speed, cognitive laziness, or a combination of these? From this intuition he proposed studying how LLM-based assistants influence users’ cognitive processes. I joined the project in 2025 through the Double Degree programme between Politecnico di Bari and Université Côte d’Azur, initially working one day per week during the first semester. The results obtained in this early phase on the experimental architecture and applications led to an offer to carry out my six-month curricular internship at INRIA Sophia Antipolis, full-time, continuing the same project as the basis for this thesis. There I joined a multidisciplinary research team: besides Professor Legout and me, who handled the technical work, it included Christopher Leturc, an expert in game theory, and Fanny Verkampt and Isabelle Milhabet, experts in psychology, whose complementary expertise was essential to design an experiment that was both technically reliable and methodologically sound from the perspective of the behavioural sciences. During the internship, we designed the studies, developed the experimental platforms, conducted live user tests, and prepared data analysis scripts. The experiments are ready to be deployed on Prolific, and the collected data will support a statistical evaluation of how LLM-powered assistants affect human reasoning, confidence, and decision-making.