“…While generation of CLA has been explored with various technical backbones, including template-based approaches [19,39], symbolic planning [56,65,81,87,98], casebased reasoning [38,83,85,91,95], or character simulation [15,61], LLM technologies powered by transformer architecture [96] have brought in a large leap in flexibility and accuracy of text generation, as these models could be "prompted" to serve arbitrary natural language tasks [12,74,94]. With these LLM capabilities, many writing tools have been introduced, and one type of tool is those that suggest text phrases to the user's writing, which is often called human-AI co-writing [2,36,53,66,103]. Researchers studied how these LLM suggestions can change people's writing and found that generated texts could spark new ideas [14,18,36], lower grammatical errors [58], and increase vocabulary diversity [58], while introducing cognitive challenge of integrating generated texts into the user's writing [14,90].…”