NSF-NSERC: HCC: Small: Interaction Techniques for Establishing and Managing Long-Term Human-AI Collaborations in Game Design and Document Writing

Harvard University, Cambridge MA

Investigators

Abstract

As artificial intelligence (AI) systems become more powerful and self-determining, they are shifting from simple tools to collaborative partners that can work with humans on complex, long-term projects. However, current AI systems may not effectively maintain communication with humans over extended periods. They often fail to track preferences, follow instructions, and/or adapt to project-specific context. The proposed work will investigate how to design AI systems that can establish and maintain a shared understanding with humans when collaborating over long-term projects. The work will advance the national interest by developing foundational principles and software for trustworthy AI systems that can support human creativity and productivity in areas such as game design, long-form document writing, and software development . The project will produce open-source tools and guidelines that will benefit industry developers, researchers, and users of AI systems across the United States and Canada. Ultimately, these AI systems will enable more effective and reliable AI assistance across the timeframes of real-world projects and accelerate the development of future AI applications. This research develops novel interaction techniques and technologies for establishing, maintaining, and managing common ground between humans and AI systems during long-term collaborative projects. Common ground represents the intersection where AI systems remain aligned with user intentions as projects evolve and grow complex. This research introduces and explores the interaction model of an “intent specification” (IS). In this work, IS means a human-readable representation of user goals, preferences, and project understanding that grounds long-term AI decision-making. The project investigates three key research questions through empirical studies and system development: 1) how to help users construct and refine their intent by measuring the effect of design decisions informed by cognitive learning theories; 2) how to align AI system understanding with user intentions through interface designs that reify grounding acts from communication theory; and 3) how to manage the accumulation of user intent and project understanding at scale efficiently and accurately through the development of semantic conflict detection and resolution techniques. Findings will be validated through two application domains, game design and long-form document writing. These domains will test generalizability of approaches with empirical methods. These methods will include usability studies of system usefulness (controlled and in-the-wild) and technical evaluations. The project will produce IntentTracker, an open-source software library for managing user intent in AI systems, along with prototype applications GameJammer and DocJammer. By combining theoretical foundations from research in cognitive science and communication theory with evaluations of interface designs and techniques, this work will create design principles for next-generation AI systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.