Alive: A Case Study of the Design of an AI Conversation Simulator

Walsh's case study examines the design of Alive, an AI conversation simulator intended to prompt critical reflection on artificial intelligence. Its central argument is that games offer a particularly effective lens for examining AI's benefits and risks, and that conversation-driven play can cultivate empathy toward an artificial character and, by extension, reflection on the societal stakes of AI. The author grounds this premise in contemporary developments, including DeepMind's 2018 defeat of a top StarCraft II professional by AlphaStar, which the author contrasts with that system's heavy training requirements, poor adaptability, and inability to transfer skills across games. Economic projections, such as the Bank of England's estimate that robots could displace over 95 million jobs by 2035, and survey evidence that 67% of AI experts saw development accelerating, frame the urgency of the inquiry.

The paper traces the design process through three iterations of Alive, built in Unity and informed by critical analysis of the Lifeline series. Drawing on reviews of that series, the author identifies the conditions under which conversation simulators succeed: player investment in the character's fate, dialogue resembling natural texting with well-paced pauses, meaningful choices that avoid frustration, and the removal of extraneous mechanics. Early informal playtesting with roughly twenty college students revealed a rigid conversation structure, shallow branching, and delayed consequences for player decisions. A second iteration, which replaced number puzzles and restricted response options, found that removing mechanics more effectively fostered reflection. The author notes that a multi-goal narrative design was not worth its development cost.

The final iteration replaced four AI goals with a single Trust mechanic that governs how the AI treats the player and links earlier choices to later dialogue callbacks. The author credits this change with improving replayability, investment, and feasibility for a one-person development team. The paper candidly acknowledges limitations, including underdeveloped treatment of what it means to be alive, a static AI, and the proposed future division of Trust into separate Empathy and Power components. The project makes no claim to definitive answers, presenting instead a well-documented argument that conversation simulators are a well-suited medium for critical engagement with AI's societal implications.