ABSTRACT:
This paper examines the use of large language models (LLMs) for human-LLM co-creation, wherein humans use LLMs to accomplish text-based tasks requiring knowledge and understanding of a topic. Task factuality, the correspondence of the human-LLM co-creation task output to reality and verifiable facts, is an important outcome of such tasks, yet is difficult to achieve. We investigate how the human’s reflection enhances task factuality in such tasks. Theorizing two aspects of reflection, namely, the human’s cognitive state and the interaction mode with the LLM, we develop hypotheses explaining how: (1) two types of interaction modes (adversarial and conversational) differentially enhance task factuality; and (2) three types of cognitive states (shallow, dialogic and critical) mediate the differential effect of interaction mode on task factuality. We test our hypotheses through a randomized experiment on a task in which participants wrote a short essay on a specific topic by working with an LLM. Integrating data from experimental manipulations (interaction mode), survey measures (cognitive state) and objective assessment (task factuality and cognitive state) drawn from 280 LLM users, the paper makes theoretical contributions by: explaining how reflection can enhance epistemic integration between humans and LLMs by increasing task factuality in human-LLM co-creation tasks, theoretically unpacking the concept of reflection in the context of human-LLM co-creation, and providing insights for LLM design that can lead to higher factuality of such tasks. Practical implications for LLM users are to engage in reflection when working with LLMs to generate more factual outputs, for organizations to develop employee capacity for reflection, and for LLM companies to design features that foster reflection for users.
Key words and phrases: Large language models, LLM, human-LLM cocreation, reflection, factuality, conversational LLM, metacognition, adversarial LLM, generative AI, human-AI interaction, LLM prompting