Words reveal wants: How well can simple LLM-based AI agents replicate people’s choices based on their social media posts
Sofie Goethals, Johannes Luther, Sandra Matz · 2025
As artificial intelligence systems take on increasingly agentic roles, they begin making decisions on behalf of users rather than merely supporting them.Consequently, it becomes crucial to understand how closely these systems can replicate human choices.In this study, we examine the extent to which digital traces of user behavior can serve as a foundation for modeling individual preferences.Specifically, we use Facebook status updates, a form of self-disclosed digital traces.Based on these digital traces, the goal is to predict users' Facebook likes across various categories (e.g., Food, Movies, Public Figures, etc.), which serve as behavioral expressions of preference.Tested over 10,000 queries, we find that most categories achieve a prediction accuracy exceeding 60%, indicating generally robust performance of the Large Language Model.These findings suggest that digital traces such as Facebook status updates can reveal meaningful patterns that allow AI systems to learn more about decisions in other contexts.