Language models as interactive systems : from game-based interaction to persuasive experiences

Large language models (LLMs) can generate fluent text on a wide range of topics in response to natural language instructions. This flexibility has led to their adoption in applications where users interact with AI systems through text, including creative tools, dialogue systems, games, and educational or persuasive technologies. In these settings, the goal is not only to generate a good response, but to support interactions that help users create, decide, learn, or change their perspective. This dissertation examines how LLMs can support interactive text-based experiences across games, persuasion, and education, and what limits their effectiveness. The dissertation first studies LLMs for interactive game design through GamePlot, an LLM-powered assistant for narrative design and playtesting. The findings show that LLMs can help designers generate ideas, revise story turns, and test narratives while leaving important creative decisions in the hands of the designer. The study also reveals limitations in narrative coherence, depth, and longer-horizon interaction. The dissertation then examines persuasive dialogue by comparing GPT-4 with human persuaders in persona-grounded lifestyle conversations. GPT-4 can be highly persuasive and is often perceived as empathetic, but it asks fewer questions than human persuaders and misses opportunities to learn about the person it is trying to persuade. This suggests that convincing arguments alone may not provide the information needed for more responsive persuasion. Finally, the dissertation studies how interaction format shapes persuasive and educational outcomes by comparing essays, chatbots, and text-based games while keeping the underlying informational content as similar as possible. Users preferred the chatbot overall and reported stronger engagement and perceived importance. The game was perceived as the least effective format, even though the game condition showed descriptively higher factual knowledge retention than the essay condition. Overall, this dissertation shows that effective LLM-based interactive systems require more than fluent text generation. Their effectiveness depends on how interactions are structured, how well systems support user understanding and human agency, and how both user experience and objective outcomes are measured.