Actionable AI: Enabling Non Experts to Understand and Configure AI Systems

Non-experts can effectively configure AI systems through hands-on interaction: The study of 22 participant pairs found that people without technical expertise were able to achieve good performance levels when configuring an AI agent in the cartpole game, suggesting that direct manipulation interfaces can make AI systems accessible to general users.

Operational understanding matters more than technical understanding: Rather than needing to comprehend how AI systems work internally, non-experts developed sufficient "operational understanding" by directly influencing the agent's behavior, which enabled them to reach their goals—an important distinction for designing user-friendly AI systems.

Actionable AI democratizes access to AI-based systems: By allowing end users to configure and influence AI agents toward their own objectives, Actionable AI provides agency to non-experts who would otherwise be locked out of controlling these systems, addressing a gap left by traditional explainable AI approaches that rely on domain expertise or stable environments.

Direct manipulation interfaces are a viable design approach: The paper demonstrates that letting users interact with and adjust AI behavior through direct manipulation—rather than requiring them to understand complex explanations—can be an effective way to bridge the gap between AI capabilities and human control in uncertain conditions.