Search Results for « Human-AI collaboration »

Abstract This chapter provides an integrative overview of computational approaches to Human–AI collaboration, highlighting the conceptual and ethical foundations necessary for effective and responsible interaction. We examine configurations of human and AI control, from human-in-the-loop to fully autonomous systems, and discuss how the nature of interaction—copresence, cooperation, and collaboration—shapes task performance. Bayesian models are presented as a formal framework for building and understanding human–AI collaboration, enabling intention recognition, shared representations, and mutual adaptation. Design implications and ethical considerations are emphasized, focusing on transparency, goal alignment, bias mitigation, adaptive mental/internal models, and the preservation of human oversight. We further explore practical applications, including legible robot motion generation for communicating intent and pedagogical reasoning for shared meaning-making in human–human collaborations. Throughout, we stress that effective collaboration requires dynamic mental/internal models, continuous learning, and alignment between human and AI goals, ensuring that AI systems enhance human decision-making while maintaining accountability and trust. By combining theoretical insights, computational models, and real-world applications, the chapter offers a foundation for students, researchers, and practitioners seeking to design human-centered, ethically responsible AI systems that support collaboration across diverse domains.

AI system memory, whether systems retain prior interactions or generate responses independently, is an underexplored factor influencing innovation in human-AI collaboration. This study examines the impact of AI system memory on innovative outcomes by comparing two AI agents: adaptive, memory-based and static, memoryless. The experimental setup included two conditions: a long sequence of 20 follow-up tasks and a short sequence of 5 tasks, which each AI agent completed. The study used cosine dissimilarity, Jaccard distance, and keyword entropy to measure innovation in AI responses. Results show that static, memoryless AI systems are more likely to generate innovative outcomes than memory-based systems, challenging the prevailing IS assumption that greater adaptation leads to more innovation. The results also show that task type moderates the relationship between AI system memory and innovative outcomes. Exploitative tasks generally yield more innovative outcomes than exploratory tasks. However, this pattern varies across metrics and memory architectures.

Abstract AI system memory, whether systems retain prior interactions or generate responses independently, is an underexplored factor influencing innovation in human-AI collaboration. This study examines the impact of AI system memory on innovative outcomes by comparing two AI agents: adaptive, memory-based and static, memoryless. The experimental setup included two conditions: a long sequence of 20 follow-up tasks and a short sequence of 5 tasks, which each AI agent completed. The study used cosine dissimilarity, Jaccard distance, and keyword entropy to measure innovation in AI responses. Results show that static, memoryless AI systems are more likely to generate innovative outcomes than memory-based systems, challenging the prevailing IS assumption that greater adaptation leads to more innovation. The results also show that task type moderates the relationship between AI system memory and innovative outcomes. Exploitative tasks generally yield more innovative outcomes than exploratory tasks. However, this pattern varies across metrics and memory architectures.

Abstract How can human–AI interaction be evaluated? This chapter provides a high-level overview of three main approaches to evaluating human–AI interaction: empirical, analytical, and model based. Empirical methods are based on measurements in controlled or naturalistic settings. Analytical methods use heuristics and walkthroughs based on best practices. Model-based methods use computational or mathematical models to estimate how a user might complete a task. Model-based methods in particular are rapidly changing due to advances in computational methods. Each approach has its scope and limitations, which we discuss in this chapter.