I propose to conceptualize "thought" as a new substrate in human-AI interaction. While there is extensive research in the NLP and ML communities on augmenting large language models (LLMs) with "thinking" capabilities, these efforts primarily focus on improving AI's reasoning performance. As an HCI researcher, I explore how enabling AI to generate and utilize thought can unlock new capabilities for interacting with humans and introduce new paradigms for human-AI interaction. I argue that this conceptualization opens up new possibilities for human-AI interaction by supporting proactive AI behavior, enabling continuous alignment with user intent, and fostering more dynamic and adaptive interaction experiences. In this paper, I articulate the conceptual foundations of thought as a substrate in human-AI interaction, demonstrate its role through system examples, and envision how this paradigm could shape the future of human-AI collaboration.
Xingyu Liu (Sat,) studied this question.