The relatively recent public release of generative artificial intelligence (AI) systems has ignited a significant leap in awareness of the capabilities of AI. In parallel, there has been a recognition of AI system limitations and the bias inherent in systems created by humans. Expectations are rising for more trustworthy, human-centered, and responsible software connecting humans to powerful systems that augment their abilities. There are decades of practice designing systems that work with, and for humans, that we can build upon to face the new challenges and opportunities brought by dynamic AI systems.

AI-enabled decision support systems have repeatedly failed in real world applications despite the underlying model operating as designed. Often this was because the system was used in an unexpected manner. Our goal is to enable better prediction of how systems will be used prior to their implementation as well as to improve existing designs, by taking human behavior into account. There are several challenges to collecting such data. Not having access to an existing prediction engine requires the simulation of such a system's behavior. This simulation must include not just the behavior of the underlying model but also the context in which the decision will be made in the real world. Additionally, collecting statistically valid samples requires that test subjects make repeated choices under slightly varied conditions. Unfortunately, in such repetitious conditions fatigue can quickly set in. Games provide us the ability to address both of these challenges by providing both systems context and narrative context. Systems context can be used to convey some or all of the information the player needs to make a decision in the game environment itself, which can help avoid the onset of fatigue. Narrative context can provide a broader environment within which the simulated system operates, adding a sense of progress, showing the effect of decisions, adding perceived social norms, and setting incentives and stakes. This broader environment can further prevent player fatigue while replicating many of the external factors that might affect choices in the real world. In this paper we describe the design of the Human-AI Decision Evaluation System (HADES), a test harness capable of interfacing with a game environment, simulating the behavior of an AI-enabled decision support system, and collecting the results of human decision making based upon such a system's predictions. Additionally, we present an analysis of data collected by HADES while interfaced with a visual novel game focused on software cyber-risk assessment.