Emergent AI-assisted discourse: a case study of a second language writer authoring with ChatGPT

1 Introduction

The release and rapid diffusion of ChatGPT has sparked wildly different views of its relationship to human writing. In light of longstanding concerns about a national writing crisis (Intersegmental Committee of the Academic Senates et al. 2002; The National Commission on Writing 2003, 2004), educators have voiced apprehensions that the teaching of writing and its use as an assessment tool might be seriously compromised. Conversely, the business sector eagerly anticipated the potential for rapidly producing written content at a reduced cost.

Historically, students in the US have grappled with mastering academic writing genres (Graham and Perin 2007). The landscape of academic writing is further complicated when considering the experiences of second language learners (Booth Olson et al. 2023). These students often navigate a dynamic interplay of cultural and social values that may not always align with mainstream academic expectations (Booth Olson et al. 2023; Bunch 2013). This potential misalignment presents an opportunity for deeper exploration of how students with diverse linguistic backgrounds approach academic writing development.

There is a growing body of work on how second language learners navigate the types of writing necessary for academic research (Curtis 2023; Rahman et al. 2023; Steiss et al. 2024; Warschauer et al. 2023). The value of ChatGPT, in offering input, feedback, or scaffolding for writing, is increasingly acknowledged. Yet, a pivotal question arises: Does relying on AI-generated texts equate to academic misconduct, or can AI-assisted writing coexist with the preservation of the human author’s distinct voice and agency?

To explore this issue, we present a case study of a student who thoroughly integrates ChatGPT into all aspects of the authoring process. First, we provide a background on using ChatGPT to facilitate writing for language learners. Next, we describe our participant, Kailing, who uses ChatGPT throughout each stage of the writing process, to write an academic proposal during her doctoral studies. We then describe how activity theory can be employed as a theoretical framework to advance our understanding of how AI tools such as ChatGPT can mediate the writing process. From there, we describe our single-subject case study design and data collection, including semi-structured interviews, analysis of writing artifacts, and GPT logs. Findings from our case study indicate that Kailing applied her expertise to dynamically interact with ChatGPT during the brainstorming, research, text generation, and revision processes. Notably, throughout this process, she retained her own sense of voice and agency. Findings such as these highlight the potential of generative AI tools, such as ChatGPT, to enhance writing for language learners in academic settings.

2 Literature review

In the past 2 years, there has been a rapid increase in studies on the integration of AI in language learning (Bin-Hady et al. 2023; Fitria 2023; Kohnke et al. 2023; Liang et al. 2023; Steiss et al. 2024; Warschauer et al. 2023). Although most findings indicate that language learners could greatly benefit from generative AI tools, they face disproportionate challenges regarding accusations of academic misconduct. Notably, GPT detectors exhibit considerable bias against these students. A study that analyzed the precision of seven widely used GPT detectors on 91 TOEFL essays and 88 US student essays found that while the detectors reliably identified essays from US students, they mistakenly labeled over half of the TOEFL essays as “AI generated,” with an average false-positive rate of 61.3 % (Liang et al. 2023). Building upon this work, Warschauer et al. (2023) explore the affordances and challenges of using AI-based tools for second language learners. They pinpoint the dilemmas faced by these learners as they engage with AI tools. To echo Liang et al. (2023), they underscore the pressure on language learners to mirror the target language. However, when L2 students utilize tools such as ChatGPT, they often face allegations of plagiarism, especially when compared to native speakers (Liang et al. 2023).

In addition to detection bias, generative AI tools, such as ChatGPT, have been shown to exhibit discriminatory bias (Kenthapadi et al. 2023). Since they are trained on large corpora of data, engineers encounter challenges when attempting to audit the data for specific biases and discriminatory elements (Bender et al. 2021). For instance, large language models have exhibited biases and stereotypes related to gender (Bolukbasi et al. 2016), disability (Hutchinson et al. 2020), and language (Lee 2023). Given these findings, it is imperative to approach the use and interpretation of outputs from generative AI tools with caution, ensuring that measures are in place to counteract and mitigate inherent biases.

There is substantial evidence indicating that ChatGPT can perpetuate bias, stereotypes, and inaccuracies. However, despite its recognized limitations, it would be a mistake to overlook its potential benefits for underperforming students, such as emerging writers, students with disabilities, and language learners. Regarding job readiness, Warschauer et al. (2023) highlight the eagerness of corporations to incorporate AI into their systems, yet students designated as English learners have limited access and opportunities to learn how to effectively utilize these tools (Warschauer et al. 2023). Denying language learners the chance to learn and employ generative AI tools can adversely impact their job readiness. A recent study found that using ChatGPT for work-related tasks in areas such as marketing and grant writing reduced the time to complete tasks by 40 % while improving quality by 20 % according to a double-blinded review of work performed. Importantly, low-performing employees improved the most (Noy and Zhang 2023). A similar study from Harvard Business School assigned strategy consultants to conceptualize and develop new product ideas in one of three conditions–without AI, with ChatGPT, and with ChatGPT and an overview of prompt engineering. Again, the two ChatGPT groups improved the most, largely due to a catch-up effect among the below-average writers who improved by 43 % in their writing tasks compared to an improvement of 17 % among above-average writers (Dell’Acqua et al. 2023). Findings such as these suggest that learning tools such as ChatGPT have the potential to level the playing field for struggling writers in the workforce.

Many are concerned that generative AI tools such as ChatGPT could boost productivity at the expense of innovation and creativity. However, empirical evidence suggests that when utilized appropriately, that may not be the case. A study investigating the influence of GenAI on creative content production found that individuals inspired by AI-generated content crafted more innovative narratives compared to those drawing from other sources (Doshi et al. 2023). This impact increased significantly among writers perceived as less creative. Such results highlight ChatGPT’s potential to equalize opportunities for emerging writers.

There are legitimate concerns that technology cannot replicate the nuances of advanced writing, tasks such as these require creativity, innovation, high-quality content, structured organization, and logically sound arguments. However, a recent study investigating ChatGPT’s role in supporting students with argumentative suggested that ChatGPT aids in content generation and streamlines the writing process, including outline preparation, content revision, proofreading, and post-writing reflection (Su et al. 2023). Leveraging generative AI tools for complex writing tasks, such as argumentative writing, is particularly beneficial for language learners, who often grapple more with the higher-order components of argumentative writing compared to their native-speaking counterparts (Booth Olson et al. 2023).

Finally, it has been long argued that human feedback is irreplaceable. However, emerging studies show generative AI tools such as ChatGPT are capable of delivering high-quality feedback for students. For example, a current investigation comparing feedback from humans to that provided by ChatGPT revealed that the practical differences between the two were minimal (Steiss et al. 2024). The consistency between human and AI-generated feedback suggests that ChatGPT could be a reliable and scalable tool, potentially bridging the feedback gap often faced by language learners due to limited human resources.

In conclusion, the rapid integration of AI into language learning has opened new horizons, offering potential benefits, especially for underperforming and L2 students. However, the challenges – ranging from detection and discriminatory biases to issues of academic integrity and the potential dilution of creativity – are undeniable. These concerns underscore the importance of a balanced and cautious approach to the application of tools such as ChatGPT. Notwithstanding its limitations, the evidence suggests that, when used judiciously, ChatGPT can be an asset, providing quality feedback, fostering creativity, and aiding in content generation. Moving forward, as the field continues to evolve, continuous evaluation and mindful implementation of such tools will be crucial to ensure that they serve to augment rather than impede the processes of language learning and advanced writing.

We now turn to the participant in our case study to better understand the conditions under which she employed ChatGPT to produce academic research.

3 Theoretical framework

Activity theory, anchored in Engeström’s (1987, 1999) work and drawing upon the Vygotsky and Cole (1978) sociocultural framework, posits that human interactions are multifaceted, influenced by numerous factors, and inherently dynamic. In aligning this theory with the writing process, we lean on Kessler’s (2020) conceptualization. Within this framework: The subject (or agent) undertakes actions with a specific goal in mind. These actions are facilitated by tools or mediators. In the realm of writing, the end product (the written text) is the goal, while various aids and resources (e.g., computer, chatbot, editing software, thesaurus) facilitate the creation of this text. The interaction is also shaped by three contextual elements: Rules provide guidelines or standards for the subject’s behavior. Community represents the broader context and stakeholders that influence the subject’s actions. Division of labor illustrates how tasks related to the primary goal are distributed among participants. In the context of academic writing, a researcher (subject) may engage with colleagues, field experts, editors, and peers during the writing and revision process. Notably, tensions might arise, such as discrepancies between established writing norms and specific assignment criteria. These tensions, however, can serve as catalysts for the learner’s developmental journey (Fujioka 2014).

We build on this theory by contextualizing it within the use of generative AI tools for writing. In our case study, the subject undertakes the goal of writing a research proposal. The end product, or written text, is mediated by ChatGPT, a generative AI tool that uses language processing to facilitate writing, among other tasks. Within this dynamic interaction between text and tool, our participant is constrained by the rules of academic writing, and the context of her broader community (i.e., the reviewers of the proposal), while engaging in the division of labor by conferring with her colleagues, peers, and scholars on how to improve her writing.

Our study is also based on the cognitive process models of writing (Flower and Hayes 1981), in which writing consists of planning, translating, and reviewing, and revising. Composition is a recursive process (Berninger et al. 1996; McCutchen 1996): writers cycle through the planning, translating, and reviewing multiple times, and these stages all interact with one another throughout the composing process (Flower and Hayes 1981).

4 Methods

5 Results

Section 5.1 below presents the results that address research question number one: How does a graduate student and second language writer in English employ ChatGPT in each phase of her academic writing process? In addressing the first research question, we used the four stages of the writing process: brainstorming, research, text production, and revising to present the results, as these stages comprehensively capture the essential phases of academic writing. This framework aligns with established models of the writing process that emphasize these sequential steps as crucial for effective composition (Flower and Hayes 1981; Scardamalia and Bereiter 1987). Employing ChatGPT at each of these stages highlights its multifaceted role in supporting both the generation of ideas and the refinement of academic texts, particularly for second language writers (Ong and Zhang 2013). By structuring the analysis around these stages, the study offers a clear and systematic exploration of the student’s interaction with GenAI tools, facilitating a deeper understanding of their influence on academic writing.

6 Discussion

In this study, we critically engage with two perspectives: the nature and practice of writing in the age of advanced AI, and the implications for academic writers, particularly those situated in contexts comparable to Kailing’s.

The findings of this study align with activity theory (Engeström 1987, 1999; Vygotsky and Cole 1978), which emphasizes the interaction between subjects, tools, and the community within a sociocultural context. Kailing’s use of ChatGPT across different phases of her writing process exemplifies how generative AI tools can mediate and enhance the writing activities of second language learners. Each stage of her writing – brainstorming, research, text production, and revising – demonstrates how ChatGPT acts as a tool that facilitates cognitive processes and supports Kailing in navigating complex academic tasks, thereby enhancing her writing proficiency and maintaining her authorial voice. This dynamic interaction highlights the tool’s role in not only generating content but also in shaping and refining the user’s ideas through iterative feedback and engagement.

Moreover, activity theory’s focus on the sociocultural context of learning and development is evident in Kailing’s interactions with her academic community (Kessler 2020; Vygotsky and Cole 1978). The rules and expectations of academic writing, as well as the division of labor among her peers and advisors, create a structured environment in which ChatGPT serves as an intermediary tool (Kessler 2020). This framework supports the notion that Kailing’s developmental journey is influenced by the interplay between her cognitive abilities, the affordances of the AI tool, and the academic standards she must meet. These interactions reveal how generative AI tools can be integrated into writing practices to support language learners, reflecting the core principles of activity theory in mediating human actions through cultural and technological artifacts.

At the core of human endeavors lies the role of tools. In examining writing through the lens of activity theory, the introduction of a new tool, such as ChatGPT, recalibrates the dynamic of human-tool interaction, necessitating a comprehensive approach that integrates human, tool, and context (Engeström 1999).

Bateson’s (1972) contemplation of the blind man and his stick offers a deeper reflection on this relationship. He posits the inseparability of the tool (the stick) from the user (the blind man), thereby highlighting the symbiotic relationship between them. Similarly, Postman’s (1993) observation on the transformation of Europe following the printing press’s advent reinforces the transformative potential of tools on societies.

Communication technologies, particularly as highlighted by Ong (1982), possess the capacity to reconfigure cognitive and societal landscapes. The inception of the printing press didn’t merely augment Europe; it reshaped it. Similarly, digital technologies don’t merely enhance our traditional literacies; they redefine them.

The essence of this transformation is evident in the arena of digital literacy. The internet’s emergence was not simply an alternative medium for traditional content delivery. Instead, for many, it heralded a paradigm shift in the art of reading, writing, and research. The challenge and opportunity were to comprehend and adapt to these new digital literacies, which include online research, content creation across multiple media, and writing for online communication to potentially large audiences.

The rise of ChatGPT reinvigorates debates around the nature of writing and the impact of technology on authorship. Does it signify a mere shift in how writing is done, or does it denote a fundamental metamorphosis of the nature of writing? Our exploration of Kailing’s engagement with ChatGPT offers valuable insights into how one adept user of technology navigated this liminal space, authoring through interaction with generative AI.

Our paper highlights questions regarding the nature of writing proficiency and sheds light on the potential of tools such as ChatGPT in academic settings. The case of Kailing serves as an illustrative microcosm of this broader landscape. As demonstrated, while Kailing utilized ChatGPT as a generative tool, she never abdicated her agency, ensuring that the content produced echoed her academic integrity and authorial voice.

This intricate balance between humans and tools, mediated by contextual rules, communities, and divisions of labor (as framed by activity theory), forms the crux of our study. While the “blind man and the stick” analogy underscores the interdependency of humans and tools, Kailing’s narrative reveals a nuanced engagement, where ChatGPT serves both as a collaborator and a tool, aiding but never replacing the human intellect.

To this end, our study highlights Kailing’s changing relationship with ChatGPT in academic writing, focusing on the tool’s strengths and limitations. While Kailing appreciated ChatGPT’s assistance in employing sentence variation, she noted recurring patterns that limited its utility. Her journey underscores the balance between generative AI support and the indispensable value of human expertise in refining and enhancing writing.

The crux of our exploration is not just how AI tools such as ChatGPT can be employed, but more critically, how they reshape the writing process, challenge traditional notions of authorship, and make us rethink the nature of writing pedagogy and assessment in academic settings. At the same time, Kailing’s continued journey underscores the relationship between AI and the human experience, underscoring that while tools such as ChatGPT can offer invaluable support, their limitations necessitate human intervention to ensure high-quality writing.

7 Conclusions

We present a case study of one participant, understanding its limitations. Kailing’s processes and results may well stem largely from her own intellect, personality, or other idiosyncratic characteristics. We do not argue that these results are universal.

However, especially at the emergence of new phenomena, small case studies can help us develop hypotheses to explore through replications and larger studies. Kailing’s example suggests new possibilities for human-AI collaboration that can be of great benefit for second language writing and writers, despite generative AI’s biases, inaccuracies, and limitations.

Further work is warranted to understand how applicable this case is to the broader field of second language writing with technology, and what kinds of pedagogy are needed to help other students learn to critically exploit this powerful new tool.