Shaping Human-AI Collaboration: Varied Scaffolding Levels in Co-writing with Language Models

Paramveer S. Dhillon,Somayeh Molaei,Jiaqi Li,Maximilian Golub,Shaochun Zheng,Lionel P. Robert

2024-02-19

Abstract:Advances in language modeling have paved the way for novel human-AI co-writing experiences. This paper explores how varying levels of scaffolding from large language models (LLMs) shape the co-writing process. Employing a within-subjects field experiment with a Latin square design, we asked participants (N=131) to respond to argumentative writing prompts under three randomly sequenced conditions: no AI assistance (control), next-sentence suggestions (low scaffolding), and next-paragraph suggestions (high scaffolding). Our findings reveal a U-shaped impact of scaffolding on writing quality and productivity (words/time). While low scaffolding did not significantly improve writing quality or productivity, high scaffolding led to significant improvements, especially benefiting non-regular writers and less tech-savvy users. No significant cognitive burden was observed while using the scaffolded writing tools, but a moderate decrease in text ownership and satisfaction was noted. Our results have broad implications for the design of AI-powered writing tools, including the need for personalized scaffolding mechanisms.

Computation and Language,Human-Computer Interaction

What problem does this paper attempt to address?

This paper explores how different levels of support (scaffolding) provided by large language models (LLMs) during collaborative writing between humans and AI can affect the writing process. Through a Latin Square design experiment, participants were asked to respond to prompts for argumentative writing under three randomized conditions: no AI assistance, sentence-level suggestions (low scaffolding), and paragraph-level suggestions (high scaffolding). The study found that the scaffolding level had a U-shaped impact on writing quality and productivity, with low scaffolding not significantly improving quality or productivity, while high scaffolding showed significant improvements for unconventional authors and technologically inexperienced users. No significant cognitive burden was observed when using the scaffolding tools, but there was a moderate decrease in perceived ownership and satisfaction. The study highlights the necessity of personalized scaffolding mechanisms in designing AI-driven writing tools to promote human satisfaction and productivity.