Claiming the research expertise on human–GenAI interaction for sociolinguistics

Kelly-Holmes's article in this Dialogue invites respondents to consider how generative AI (GenAI) will impact our work in sociolinguistics. As someone whose research is primarily situated in science education, this question prompted me to reflect on the very nature of sociolinguistics and its connection with GenAI. From an outsider's perspective, I found myself asking more fundamental questions, such as, what is sociolinguistics and how does the community of sociolinguists interact with the development and application of GenAI? At the end of my reflection, I conclude that the diverse research within sociolinguistics places sociolinguists in a unique position to claim a disciplinary expertise in this emerging research, on an equal footing with computer engineers, data scientists, and technology designers often associated as the domain experts. Viewed in this light, I find Kelly-Holmes's question somewhat one-sided and reactive to the development of GenAI. Instead of framing the question as how GenAI will change our work in sociolinguistics, I like to propose a bolder and reciprocal proposition whereby our work in sociolinguistics can, will, and should change the way GenAI is being developed and applied in various social contexts. In this commentary, I will elaborate my reasoning. As a qualitative researcher, it is first necessary to briefly explain my positionality. With a background in theoretical physics and experience as a former physics teacher, my research revolves around the education of science learners in schools. My connection to sociolinguistics began during my doctoral studies at the University of Michigan when I was working with Jay Lemke to examine the role of language and multimodality in science education. My research on language issues spans across a range of topics including classroom discourse, scientific literacy, reading and writing science, and translanguaging. In addressing these issues, I learned and employed various discourse analysis approaches informed by systemic functional linguistics (Halliday, 1994) and social semiotics (Kress, 2010; Lemke, 1990). When ChatGPT was released to the public in 2022, I became very interested in researching the use of GenAI, particularly because I saw its potential in enhancing student learning through language (Tang, 2024). Upon reflection, I believe it was my training in sociolinguistics that enabled me to see unique perspectives on human–AI interaction across various aspects of social life. In the social sciences, a major problem among researchers is the tendency to view GenAI as a “blackbox” machine that simply responds to queries without unpacking the process of how it learns and generates text. In a nutshell, GenAI utilizes a large language model (LLM) that is able to learn the intricate patterns of words and phrases in natural language from a massive corpus of text data. In the training process, the model assigns probabilities to how often words and phrases co-occur in the corpus. The model then uses this probability distribution to predict and generate the most likely sequence of words in response to a user's text input. A familiarity with linguistic concepts like collocation, semantics, pragmatics, register, and genre is useful to understand how an LLM assigns probabilities to potential words and phrases that align with human language use patterns. This machine learning process is not limited to verbal data; it is also similar with data involving other semiotic modes like images, sound, and video through the use of large multimodal models or LMMs (for elaboration, see Tang & Cooper, 2024). Kelly-Holmes's article touches on this ability of GenAI when she describes AI as “a massive human language corpus” shaped by years of human interaction. She also raises crucial questions such as “What is social language?” and “Where do we locate the socio?” in AI-generated texts. I want to expand on this idea further by using Bakthin's concepts of dialogism and heteroglossia to understand how an LLM acts as an intermediary across multiple social worlds (see Tang, 2024). According to Bakhtin (1981), language use is inherently dialogic, incorporating a multitude of voices through intertextual links in the “chain of speech communication” between individuals (p. 89). Building on this idea, we can understand the AI's voice not as its own, but as dialogic and heteroglossic in two major ways. First, as LLMs draw on data from diverse sources, they juxtapose these sources to produce language that represents a complex amalgamation of human voices. Second, the interaction between users and LLMs is a dynamic co-construction of knowledge. Consistent with Bakhtin's view that no utterance or written text exists in isolation outside a chain of texts (or intertexts), the output from GenAI is the product of a continuous interplay between the user's input and the AI's data-derived responses. This Bakhtinian perspective of human–AI dialogue provides a useful framework for understanding the nature of social language and how it is distributed in our interactions with LLMs. For Bakhtin, the socio is constituted within the very fabric of language itself, as every utterance shapes and reflects the diverse voices, ideologies, and knowledge structures of various sociocultural groups. In the context of human–AI interaction, we can argue that AI-generated text is a form of social language that emerges from the interplay between the voices embedded in the linguistic data and the users. The heterogeneity of language in AI-generated text is thus a manifestation of the social diversity encoded within the vast corpus of data drawn upon by an LLM as well as our interactions with GenAI. However, unlike human-to-human dialogue, GenAI presents several challenges that could impede the heterogeneity of language and diversity of voices in human–AI interactions. First, users can control the voices manifested in GenAI by specifying the register and genre of its output. This technological feature can be a double-edge sword. While it allows users to steer the dialogue toward more tailored and detailed responses, it can also limit the dialogue by potentially excluding other perspectives. Second, the machine learning and feedback mechanisms in GenAI can introduce bias in the data when the technology is predominantly used by speakers from dominant linguistic and cultural groups. This could result in GenAI becoming more proficient and skewed toward certain voices while marginalizing others. Finally, we must remember that GenAI is developed by companies with vested socioeconomic interests. These companies have significant voices in shaping how their LLMs operate. A Bakhtinian view therefore not only highlights the inherent diversity of social language in human–AI interaction but also underscores the importance of critically engaging with the ways in which AI technologies might amplify or constrain this heterogeneity in our interactions. Applying Bakhtin's dialogism and heteroglossia to the analysis of human–AI dialogue may raise concerns about anthropomorphizing GenAI as a human speaker. After all, GenAI lacks intentionality, consciousness, and social awareness. To address this issue, I draw on a posthumanist perspective of language informed by recent developments in new materialism (Pennycook, 2017; Wee, 2021). This emerging paradigm challenges anthropocentric assumptions in the social sciences by deconstructing established binaries between humans and non-humans. Using the concept of intra-action from new materialism (Barad, 2007), I argue the importance of recognizing the shared agency between humans and non-humans in human–AI interactions, as well as the co-creation of scientific knowledge as a partnership involving both humans and GenAI (see Tang & Cooper, 2024). Building on this idea, we need to expand our view of a social language to a “sociomaterial assemblage” of language where symbolic signs and material objects are always entangled (Pennycook, 2017). In this sense, the entanglement between humans and GenAI in language use is not a radical departure from the historical relationship between language and technology through different epochs, from clay tablets to ink and paper, the printing press, digital screens, and now with artificial intelligence. Having elaborated my theoretical views, I would now like to return to my earlier point that the diverse perspectives and interdisciplinary research in sociolinguistics offer sociolinguists a unique position to shape the development and application of GenAI in society. My work as a science education researcher—an area rarely associated with sociolinguistics—demonstrates how the diversity within this field contributes to a broader understanding of language use in various social contexts. While my research focuses on the centrality of language (including multimodal representations) in science classrooms, it also connects with other sociolinguistic studies that explore occupational languages, language policies, gender studies, and multilingualism. What unites these diverse areas is a common focus on how language operates within and across social contexts in shaping and reflecting societal dynamics. Despite varied disciplinary backgrounds, sociolinguists bring a deep understanding of how language both shapes and is shaped by the sociomaterial world. It is this breadth of research that positions sociolinguistics to offer critical insights into the use of GenAI, which metaphorically speaking, is better regarded as a massive language mediator than a technological wonder. From a Bakhtinian perspective, sociolinguists are well-positioned to inform the research and development of GenAI. Human–AI interaction, as mediated through LLMs and LMMs, mirrors the heteroglossic nature of language itself, where AI-generated text reflects the complex amalgamation of voices and perspectives within its training data and interaction with users. The broad spectrum of sociolinguistic research enables the field to investigate the manifestation of different voices in various social situations and advocate for the inclusion of marginalized voices often sidelined in our interaction with GenAI. The challenge and opportunity for sociolinguists lie in ensuring that our use of GenAI in social contexts accurately represents and incorporates the wide range of linguistic and cultural perspectives that exist in society, particularly those of underrepresented groups. For instance, in a recent study conducted in collaboration with a high school in Western Australia, I applied Bakhtin's ideas to explore the integration of GenAI into classroom discourse through a dialogic approach (see Tang et al., 2024). The study views GenAI not as an authoritative knowledge provider but as a dialogic partner that facilitates collaborative dialogue and co-construction of knowledge. This approach encourages students to critically engage with AI-generated content and integrate multiple perspectives into their learning. The study also demonstrates how human–AI interaction was analyzed from a sociolinguistic perspective by tracking how the diverse voices present in the discourse evolved through the interaction. This example underscores the importance of understanding and maintaining the heteroglossic nature of language in GenAI, thereby ensuring that the technology does not merely reflect dominant cultural norms but rather encourages the inclusion of diverse voices in the learning process. By fostering an environment where students interact with GenAI as a dialogic partner, my research highlights how sociolinguistic principles can guide the design and deployment of GenAI systems to amplify rather than diminish the diversity of human communication. In sum, the diverse and interdisciplinary nature of sociolinguistic inquiry—from the analysis of classroom discourse to the impact of language policy and multilingualism—is a strength that allows sociolinguistics to claim a research expertise on how GenAI interacts with various social worlds in education, government, healthcare, business, the arts and humanities, and the sciences. The insights generated by sociolinguists provide much-needed perspectives that balance the often technocentric views from computer engineers, data scientists, and technology designers, who are traditionally seen as the primary experts in GenAI. By applying sociolinguistic principles into the development and analysis of these technologies, we can make a difference to the evolution of GenAI in a way that is not only technologically efficient, but also socially responsible and inclusive. Sociolinguists, therefore, have a crucial role to play in shaping the future of human–GenAI interaction to reflect and uphold the diverse and dynamic nature of language and communication. The author declares no conflicts of interest. No funding was received to assist with the preparation of this manuscript.