The Imitation Game: Evaluating Persona-Driven LLM Response Behavior in Web Surveys

Persona conditioning prompts LLMs to adopt demographic profiles (education, gender, generation, stance) to approximate human variation. We analyse how such persona-conditioned models behave synthetic survey data generation, quantifying persona effects, applying multi-layer community detection to assess cross-question consistency, and comparing linguistic and sentiment patterns between human and LLM text. Overall, persona conditioning amplifies identity cues while reducing internal diversity, yielding coherent but less human-like variation. We discuss implications for survey realism and provide diagnostic methods for detecting and analysing synthetic respondents.