Exploring the Use of Generative AI Chatbots for Second Language Writing Feedback with Japanese University Learners of English as a Foreign Language

Abstract

The emergence of generative AI (genAI) chatbots like ChatGPT presents both opportunities and challenges for second language teaching and learning.This is especially the case for writing due to AIs ability to quickly produce natural human-like texts by means of simple prompts.The present study explored the use of genAI chatbots not as composing but as feedback tools used by students on texts they wrote for an oral presentation.Participating students (N6) answered post-task questionnaires and submitted their draft and revised texts as well as chat logs for analysis.Results indicated that genAI could be used cautiously as a source of reference in addition to instructor feedback.In light of use-and feedback output-related issues raised, the need to develop AI literacy and better understanding of genAI through further research are highlighted.Exploring the Use of Generative AI Chatbots for Second Language Writing Feedback with Japanese University Learners of English as a Foreign Language (WHITE) - -These abilities were based on the GPT (generative pre-trained transformer) large language model (LLM) GPT-3, an artificial neural network built with human feedback-reinforced machine learning using a large and diverse dataset of human language (first available via API in 2020; OpenAI, 2020).While computer text generation of individual sentences based on predictive algorithms had already existed for some time (familiar to most people now through text-completion suggestions of mobile devices and email and word-processing programs), the transformer model incorporated by OpenAI into its GPT family of models allowed for greater "attention" and context awareness to not only construct individual grammatically correct sentences but entire natural-sounding, human-like texts with a high degree of rhetorical consistency (Godwin-Jones, 2022, 2023).In effect, it appeared to be able to not only produce texts but also, because of its dialogic interface, communicate with users based on the messages (prompts) they entered.The fact that this ability was based on large amounts of information on the internet in its training data ( including data from CommonCrawl, book corpora, and Wikipedia; Brown et al., 2020;Zhao et al., 2023) meant thatChatGPT was able to provide clear (if not always accurate) answers to user questions typically found through more involved internet searches, thus giving it an air of authority and perceived useability/productivity of a "super tool" with potential revolutionary uses.However, ChatGPT and similar generative artificial intelligence (genAI) chatbots are fundamentally language-processing rather than information-processing tools (Walters & Wilder, 2023).As a result, the limitations of the technology became apparent: information was not up-to-date and limited to the cutoff date of the training data, and even then, due to its "emergent" nature it could "make things up" or "hallucinate" information and ideas in achieving its intended purpose of completing natural-sounding human language texts (OpenAI, 2022).Initially having a training cutoff date of September 2021 for the first-released and initial free version, , the paid version, ChatGPT Plus, provided access to an updated model, GPT-4, with a more recent December 2023 training cutoff date (OpenAI, n.d.-b).GPT-4 also showed increased information accuracy and reduced instances of hallucinations (OpenAI, n.d.-a) and remains the most advanced model reserved for ChatGPT Plus users while free and paid users both have access to the better-optimized GPT-4o, released May 2024 (OpenAI, 2024).GPT-4 and GPT-4o have a number of other features, including multimodal (not just text) input, including file uploads, and can, via modifications first through plugins and now through custom instruction "GPT"s, connect to the internet for more recent information and services.The pace and degree of technological change mean that we will likely see continued and rapid improvement (Godwin-Jones, 2022).The emergence of ChatGPT and similar LLM-based chatbots such as Anthropics Claude, Googles Gemini (formerly Bard), and services that use OpenAIs GPT models such as Microsofts Copilot (formerly Bing Chat) have generated a large degree of excitement for their potential applicationsboth positive and negative.ChatGPTs ability to generate not only text, but also images, computer code, and results of data analysis, all multimodally, have

Ähnliche Arbeiten