AI, power and sociolinguistics

Ico Maly is associate professor Digital Culture Studies (Tilburg University, The Netherlands).

In her opening essay, Hellen Kelly-Holmes asks herself and us ‘how Artificial intelligence will change the way that sociolinguists carry out research’. Instead of giving a clear-cut answer to that question, I would like to take one step back. Before we can think about the concrete ways sociolinguists can use artificial intelligence (AI), it would not be a luxury to first have a sociolinguistic theory on AI. AI is not a neutral tool, it has its own epistemology, produces specific discourses and changes sociolinguistic environments. I do not pretend to have such a full-blown sociolinguistic theory of AI, but I would like to use this opportunity to give a first preliminary sketch of what such a sociolinguistic theorization of AI could look like.

Starting with the latter, it strikes me how Kelly-Holmes downplays her own work and states that ‘the writing (of ChatGPT) is substantially more correct than my own rambling’ (Kelly-Holmes, 2024). She is clearly not alone in such an assessment of AI. Most users of ChatGPT are equally impressed. It explains the success of the app among our students, and the world at large. By February 2023, the app had 100 million people using it on a weekly basis. And in 2024, that number would rise to 180 million. ChatGPT is now so omnipresent that we have to understand it as a cultural force.

The discourses ChatGPT produces are being used in a vast number of fields: journalism, law, academia, marketing, politics and digital culture in general. And more, the app is now also embedded in social media like Instagram. Other companies have their own LLMs implemented in search engines, smartphones and social media platforms. AI generates language and is used to moderate language, to help you search, to give you a more personalized digital experience and much more. AI has become a central social structure (re)producing and policing language. And in that sense it gives direction to discourse and culture.

It is exactly this success that warrants sociolinguistic attention as it has effects on individuals, society and language. On the most micro-level, understanding the relation between AI-produced language and society warrants studying it as interaction. When we do that, we see that users are entering a specific type of communicative relation with specific communicative norms. One entity—the human—is taking up the role of the one asking for information, placing the other—the AI—system in a position of knowledge. This framing of the AI bot as the producer of knowledge is a cultural format. It is steered by the example prompts on the ChatGPT website, but also by the many social media pages and YouTube videos that are dedicated to developing the ‘correct prompts’. The other side of the interaction—the chatbot—is programmed to respond in particular ways. This specifically programmed relation is inherent in the design of the chatbot—think of ChatGPT slowly typing text—and is based on what ‘human-like’ language looks and feels like (Hicks et al., 2024). Gershon (2023) rightfully calls ChatGPT a ‘genre producer’. The text that we as users see being produced by ChatGPT is not solely the result of how the model ‘learns’ by itself. It is also shaped by human feedback (alignment and super-alignment in AI jargon) that helps reduce some of its major limitations (Lenci, 2023) and produce recognizable genres.

Hicks et al. (2024) subtly argue that ChatGPT is a bullshit machine that is only designed to convey an impression of human-like language and thus not designed to produce truth, knowledge or understanding. LLMs reproduce a statistical ‘common sense’ in relation to the prompts they receive. Hick et al.’s claim that ChatGPT is bullshitting refers exactly to this specific epistemology of LLMs. Those models, they argue, are not designed ‘to represent the world at all, instead they are designed to convey convincing lines of texts’ (Hicks et al., 2024, p. 38). From this perspective on LLMs, their so-called hallucinations are not exceptions; on the contrary, they are indexes of this particular epistemology of LLMs. LLMs are not build to produce truth, but to convince us (users and investors) that they are performing in a performant, maybe even magical way. As a mythical super-computer talking to us, mimicking a human.

Whether this bullshitting works depends on us, humans. In the end, it is still this human that needs to assign meaning to ChatGPT's answers. The trust that many ChatGPT users attribute to the ‘language’ it produces shows that ChatGPT performs very well. It does succeed in building trust. Users seem to be impressed by the language it produces. So much so, that they use it as ‘knowledge’. Assigning epistemic authority to ChatGPT's responses entails that those responses will have a meaningful impact and effect. One that will be re-produced through re-entextualizations. This ‘epistemic trust’ forms the foundation of a very specific power relation. Analyzing this power relation is exactly what sociolinguistics should commit to. Not only because it is academically relevant, but because it is societally important too. It reshapes language and meaning-making processes in society and thus cultures and societies in general.

Analyzing the discursive power of ChatGPT from an interactional point of view means analyzing the questions and media-ideologies of the users in relation to ChatGPT's answers and how people work with those answers. I, of course, do not have access to the interactional history between Kelly-Holmes and ChatGPT. What I do know is that ‘my ChatGPT’ answers the same research question Kelly-Holmes asked ChatGPT very differently. No matter how many times I ask the same question, with different logins, I never get the exact same answer as Kelly-Holmes. Especially the form is different: I get bullet points summing up what AI will do for sociolinguistics. Interestingly, what is quite similar in all the answers Kelly-Holmes and I received from ChatGPT is the technological solutionism. In Kelly-Holmes's case, five out of six paragraphs stress how AI could be of help to sociolinguists. And in my case, ChatGPT provided nine categories in which AI can change sociolinguistics. There seems to be no doubt in ChatGPT's ‘mind’ that AI is a blessing for sociolinguistics besides some smaller ethical considerations and the warning that ‘the importance of human interpretation and context cannot be understated’ (Kelly-Holmes, 2024).

In the last months, I asked ChatGPT tons of questions for my own research (mostly related to controversial topics), and I found this ‘both-side-ism’ a consistent feature of ChatGPT's understanding of knowledge. It is a format that, especially in the public sphere, connotes ‘scientific knowledge’ and is thus very powerful and convincing. Interestingly, this format is also very powerful in what it doesn't address. What was completely absent in ChatGPT's responses to the queries of Kelly-Holmes and myself is a true understanding of an actual ethnographic approach to sociolinguistics and language. An approach that is informed by a very different epistemology and ontology than the one fuelling the current LLMs. Language, from an ethnographic perspective, is understood as language-in-use, as language mobilized by humans in specific layered, polycentric and stratified contexts. While language from an ethnographic perspective is always inextricably linked to the world, LLMs do not know the world. Even more, the fact that LLM's approaches and ethnographic approaches to knowledge are not compatible is glossed over by ChatGPT. That becomes all the more obvious when ChatGPT suggests to Kelly-Homes that AI can be used ‘to conduct digital ethnography by examining online communities, forums and virtual spaces’.

ChatGPT's suggestions are convincing on a surface level, but from the moment we start scratching the surface, we encounter meaning problems. Sociolinguists bring with them a very specific ontology and epistemology. They see language as ‘socially loaded and assessed tools for humans, the finality of which is to enable humans to perform as social beings. Language, in this tradition, is defined as a resource to be used, deployed and exploited by human beings in social life and is hence socially consequential for humans’ (Blommaert, 2018, p. 4). Humans learn language in specific cultural contexts, and it is those contexts that determine not only ‘what is available’ but also that ‘what is available is laden with the meanings of that culture’ (Kress, 1997). Language is here understood as the architecture of society, the resources by which people create meaning. As Kelly-Holmes rightly stresses, this sociolinguistic context is missing in LLMs. LLMs are not programmed to ‘understand’ language, but to generate ‘language’ on the basis of statistical calculations (Lenci, 2023). Transformer models, argue AI programmers (Meritt, 2022), allow AI to understand ‘context’. These transformer models explain the massive step forward in language models since 2017. Those models take as ‘input a whole sentence or discourse and generate a contextual embedding for each of its word tokens’ (Lenci, 2023), which means that one word can be assigned two or more different ‘contextual’ meanings depending on the model. Even though this is a huge step forward for AI, it is still very far away from how sociolinguists and ethnographers understand ‘context’. In Silverstein's terms, we could say that transformer-based foundational models allow LLMs to have a statistical proxy for the ‘understanding’ of a denotational text in terms of ‘co-textual structuring of signs’ (Silverstein, 2023, p. 33). Even though that is a breakthrough, it doesn't mean that LLMs ‘understand’ language and context in the same way as humans. The latter use all kinds of real-life experiences to make sense of text and attribute identity characteristics and moral judgements to discourse based on specific usage of words in specific contexts. That is still completely absent in LLMs.

The epistemology of AI is different from the way language is learnt by humans, and the absence of any cultural understanding is key to this. Language in LLMs is about statistical co-occurrence, not about deep cultural meanings. As a result of this specific LLM ontology and epistemology, the data used are of crucial importance. LLMs are trained on large internet-based datasets, but large doesn't mean equal or diverse. On the contrary, it is known that those databases are heavily biased, and that is exactly why those models need ‘alignment’ in the form of bias removal. Bender et al. (2021, p. 614) argues that from ‘each step, from initial participation in Internet fora, to continued presence there, to the collection and finally the filtering of training data, current practice privileges the hegemonic viewpoint’. Combine this with a rising trust in the authority and knowledgeability of those LLMs and we see where sociolinguistics can come in.

We now understand that the issues raised by Kelly-Holmes are much more important than the bullet points ChatGPT serves us. If we think about the invisibilization of mediation and technology in the production of discourse, a topic Kelly-Holmes rightfully raises, we are not only encountering a problem for research, it is also deeply connected to the reproduction of power. What is invisible cannot be questioned. And that is exactly what sociolinguists should do, find ways to question the cultural power of AI. And that means that we need to find ways to ‘make visible’ the cultural impact of AI produced and policed discourse. In essence, it forces us to analyze AI in terms of which voices it reproduces and which voices remain invisible. ‘Digital language’ is not part of a different domain, and shouldn't be studied as such. It already reshapes society.

To conclude this reflection, I would like to come back to the main question Kelly Holmes asked ChatGPT: how artificial intelligence will change the way that sociolinguists carry out research and highlight what stays invisible in ChatGPT's answers. Is it not remarkable that ChatGPT explicitly focuses on how sociolinguists could use ‘AI’ in their research, but not once addresses how AI reshapes the field we study? In other words, not only does ChatGPT normalize the use of AI by sociolinguists, it also obfuscates how AI functions as a digital ideological infrastructure that co-constructs language that in turn co-organizes society. Before using AI as a research tool, we should first critically analyze AI and build towards a sociolinguistic theory of AI. I agree with Kelly-Holmes that sociolinguists should above all focus on making this ‘invisible’ impact visible. More than asking what is ‘real’ and ‘authentic language’, we should focus on the function and impact of ‘AI (re)-produced language’ in context. Not real versus not real, or authentic versus non-authentic should be our object of research, but how people use AI-produced language in society. That entails focusing on how humans use and interact with AI-produced language, the meaning they assign to it and what this says about the ideological and cultural power of AI.

The author declares no conflicts of interest.