Large Language Models (LLMs) have been increasingly integrated into a variety of tasks, facilitating human endeavors in generating creative outputs, ranging from product ideation to digital artwork. Such novel capabilities of LLMs have ushered in a new era of collaboration between humans and Artificial Intelligence (AI), which has grabbed the attention of researchers and practitioners alike. Thus, in this dissertation, I explore the intersection of emerging LLMs and creativity, with a primary focus on writing tasks. This dissertation includes two studies. In the first study, I examine the impact on perceived creativity of varying levels of generative capabilities of LLMs - namely, randomness, which has been overlooked so far and which is manipulated via a quasi-experiment. I find that collaborating with an LLM with high randomness that generates more diverse advice does not necessarily lead to increased perceived creativity of work, as the role of humans matters. Moreover, I explore how the characteristics of human evaluators and their perceived extent of AI use influence their assessments of creativity. In the second study, I focus on growing concerns regarding the potential misuse of generative AI, particularly its capacity to produce plagiarized content. Motivated by the divergent thinking creativity literature using the Divergent Association Task (DAT), I construct DAT(Sent), a metric to proxy semantic dissimilarities within a document, and further propose an effective GPT detector classifier, GPT-DATector. I show that on average, human-generated contents have a larger DAT(Sent) than AI-generated texts across different writing tasks and datasets. Empirical evaluations demonstrate that the proposed GPT-DATector outperforms state-of-the-art models in terms of prediction performance. Most importantly, GPT-DATector has the potential to reduce bias in the detection of AI-generated text.
Read
