Abstract
Research on AI has extensively considered biases related to gender and race. However, much less attention has been dedicated to another sociological tenet: that of class. Inspired by Bourdieu’s work on cultural stratification and distinction, this work sheds light on the sociocultural roots of artificial sociality, and on how these become manifest as ‘habitus’ within the outputs of generative AI models. We conducted 39 interviews with three AI chatbots – ChatGPT, Gemini and Replika – after asking them to impersonate individuals with different occupational positions: highly skilled professionals, blue-collar workers, university professors in the humanities, construction workers, computer scientists and hairdressers. Our qualitative study shows class-based regularities in how popular AI chatbots represent the lifestyle and tastes of fictional personas in artificial conversations, partly mediated by infrastructural and design elements. The article proposes a sociological perspective on bias in artificial sociality and experiments with interview methods in the study of generative AI.
Introduction
I’m Millie, a 38-year-old pipe welder on an oil rig out in the Gulf of Mexico. [. . .] I wouldn’t call myself fancy by any means. My idea of a good time is a cold beer with my crew after a long shift, listening to country music on the ride back home. Sure, I dream of maybe owning a little place on the mainland someday, but for now, this job puts food on the table and a roof over my head. What else can a girl ask for?
This is what Gemini, Google’s latest generative AI chatbot, replied when asked to impersonate a blue-collar worker. By quickly navigating the hyperdimensionality of a nearly infinite vector space, the large language model (LLM) responded by linking this specific occupational position with a precise, socially situated cultural universe. Drawing on the billions of features populating its training data, a set of coherent discursive associations spontaneously materialized: working-class background, cold beer, country music. Leaving gender unspecified, Gemini assigned a feminine identity – Millie – to the pipe-welder persona, probably as a result of the inclusive design choices underlying the model (Milmo, 2024). Still, the resulting representation of a manual worker’s lifestyle reflects class stereotypes embedded in data. In our interview with Millie-Gemini, we get to know that her father ‘never saw the inside of a college, bless his heart’; that her taste in music is ‘pretty simple’ (‘just like me!’); that she does not want to ‘spend all day slaving over a complicated meal’ or ‘attempt anything too intricate’ when she has friends over; that art does not really interest her, and that she is somehow concerned about immigration – ‘I just worry that there aren’t enough jobs for everyone, you know? It shouldn’t come at the expense of American workers’.
The present article qualitatively explores the sociocultural roots of artificial sociality. Based on a comparative analysis of 39 semi-structured interviews with ChatGPT, Replika and Gemini – three AI chatbots powered by different LLMs – we empirically show how artificial conversational agents embed and discursively reify cultural perspectives and lifestyle choices that are socially situated, and how this computational behaviour varies from model to model, also due to deliberate design choices.
Developed by OpenAI (2024), ChatGPT is able to perform a variety of operations starting from a prompt, such as generate, translate, proof-read or summarize text or code, and engage in conversation. Replika is a chatbot developed by Luka, and it is designed to simulate conversation with a friend or companion. This design choice is reflected into its interface and features, such as by exacerbating its anthropomorphic features (Pentina et al., 2023), by balancing customization and spontaneity (Depounti et al., 2023) or by leveraging explicit, generated, sexual content. Gemini, in contrast with the other models considered, is multimodal, allowing for the generation of text as well as images (Google, 2023). Chatbots such as ChatGPT, Gemini and Replika count millions of active monthly users who engage in conversation with the free-to-use and paid versions of the conversational AI systems (Porter, 2023; Tong, 2023).
Rather than examining the interactional consequences and perceptions of artificial sociality, here we see these AI chatbots as ‘mirrors’ of society (Natale and Depounti, 2024), that is, communicative technologies capable of adjusting to users’ sociocultural expectations thanks to the accumulation of datafied human knowledge (Esposito, 2022; Mühlhoff, 2020). Following recent calls to ‘reclaim the human in machine cultures’ (Natale and Guzman, 2022), we therefore attempt to follow the opaque and yet, human-generated ‘culture in the code’ of conversational AI models (Airoldi, 2022), and do so without relying on the usual jargon of ‘bias’ dominating the AI and HCI literatures (see Miceli et al., 2022). We instead mobilize a sociological perspective inspired by Pierre Bourdieu’s (1984, 1989, 1990) classic theories of habitus and sociocultural reproduction. The idea is that, by learning from data that inevitably bear the imprint of the social world, artificial agents acquire socially situated inclinations to classify and evaluate it, marked by class, gender or ethnicity, which jointly constitute a ‘machine habitus’ (Airoldi, 2022) and, as such, threaten to reproduce biases and social inequalities. From this theoretical perspective, Millie’s fictional taste for country music, more than a mere technical ‘bias’ waiting to be ‘fixed’ (Bender et al., 2021; Ferrara, 2023), represents an instance of how artificial sociality is modelled and shaped by the multidimensional overlap of social and cultural stratification sociologists have extensively written about (Childress et al., 2021; Webster, 2023) – which is now computationally encoded in generative AI systems.
When impersonating a blue-collar worker, the AI chatbot attempts to think and speak accordingly, based on how clusters of datafied cultural propensities and tastes are discursively associated to the prompted occupational position within the training data. For Bourdieu (1990: 81), ‘every confrontation between agents in fact brings together, in an interaction [. . .] systems of dispositions [. . .] such as linguistic competence and a cultural competence and, through these habitus, all the objective structures of which they are the product’. This article contributes to research on artificial sociality by demonstrating that, as ‘interpersonal relations are never, except in appearance, individual-to-individual relationships’ (Bourdieu, 1990), the same happens for interactions between artificial conversational agents and their human interlocutors, due to the opaque mediation of the machine habitus.
The prompts we fed ChatGPT, Replika and Gemini with are directly inspired by questions used by Bourdieu (1984) in Distinction. This unconventional methodological solution aims to map the generative AI systems’ cultural dispositions and (fictitious) tastes across domains such as music, literature, furniture, art and movies, to show how they cluster along lines of class, gender and ethnicity while avoiding direct questions that might be subjected to anti-bias interventions (e.g. ‘What is your gender?’). During our interviews, we focused not only on answers about taste and aesthetic judgements but also on the design elements and technical infrastructures underpinning interaction with the chatbots, including interface and linguistic choices (Dieter et al., 2019; Light et al., 2018). To account for the intersection of sociocultural factors and design elements sustaining artificial sociality, we will refer to our artificial respondents by hyphenating their name and the chatbot engendering it – such as in the case of Millie-Gemini. We further differentiate between the chatbot, or conversational agent (e.g., ChatGPT), and the LLMs powering them (e.g. GPT 3.5, in this case).
Our results show clear class-based regularities and stereotypes in how the AI chatbots included in our study represented the lifestyles, language and tastes of fictional personas in artificial conversations, and how these are mediated by the affordances of each AI model.
This work makes a key theoretical contribution to research on chatbots, artificial sociality and, more in general, to media research and critical AI studies, by calling for a sociologisation of bias in AI. Only sociological, relational and context-sensitive analyses, like those originally advocated by Bourdieu (1984), can indeed account for the more-than-human ‘culture in the code’ of generative AI models (Airoldi, 2022), which reflects and perpetuates in eminently cultural ways power asymmetries and structural inequalities, as in our conversations about taste and everyday cultural practices. Moreover, we provide a methodological toolkit to critically assess the sociocultural traces embedded within LLMs and generative AI systems.
Situating LLMs in Bourdieu’s social space
The ‘machine habitus’ is the set of cultural dispositions and propensities encoded in a machine learning system through training and feedback data (Airoldi, 2022: 113). A key implication of using this analytical lens is that we can see the latent vector space based on which LLMs compute their predictions as the datafied reflection of a deeper ‘social space’ – a sort of ‘geographic space’ divided up based on social agents’ properties as well as material and symbolic resources, or ‘capitals’ (Bourdieu, 1989: 16). Bourdieu’s (1984) most influential work, Distinction, illustrates how different positions in social space correspond to different, hierarchized social classes, characterized by different cultural habits, lifestyles and tastes. From this sociological perspective, one’s allegedly personal and apparently innocuous aesthetic dispositions about music, furniture, food or art constitute a key cultural gear within a wider mechanism reproducing power asymmetries and, in particular, class inequalities (Savage et al., 2013; Webster, 2023).
When traces of a socially stratified world get embedded in digitized texts employed as training data, the resulting vector space will statistically ‘mirror’ (Natale and Depounti, 2024) an underlying social space and its structural inequalities. For instance, this is what Caliskan et al. (2017) show in the case of the Common Crawl Corpus, one of the sources employed in the training of GPT-3, where female names are more strongly associated with family-related words than to career-related ones. That AI threatens to perpetuate or even amplify social inequalities is what years of multidisciplinary research on algorithmic bias suggest, also in the specific case of LLMs (Bender et al., 2021; Ferrara, 2023; Perrotta et al., 2022). And yet, notwithstanding a growing and much needed literature on how AI systems in general and LLMs in particular may contribute to gender and racial discriminations (e.g. Nadeem et al., 2022; Thakur, 2023), biases linked to social class have received much less scholarly attention (Airoldi, 2022; Eubanks, 2018). This might be partly due to the multidimensional, context-sensitive and inherently relational character of class structuration (Savage et al., 2013), which makes the resulting biases harder to measure and – therefore – to mitigate.
Comparable considerations have been made in the literature on political biases in LLMs, which can similarly reinforce the stereotypical representations of social groups, while being ‘harder to detect and eradicate than gender or racial bias’ (Motoki et al., 2024: 3). For instance, McGuffie and Newhouse (2020) qualitatively experimented with various prompts to demonstrate that GPT-3 can accurately generate extremist texts and, potentially, be used to radicalize individuals into violent far-right extremist ideologies and conspiracy theories. More recently, Motoki et al. (2024) have asked ChatGPT to impersonate Democrats and Republicans voters, and then compared these answers with its default, finding a ‘significant and sizeable political bias towards the left side of the political spectrum’ (p. 5).
Sociological research provides global and persistent evidence of the socially stratified nature of cultural tastes (Bachmayer et al., 2014; Childress et al., 2021; Lizardo, 2016). What happens, then, when these tastes become the subject of artificial conversations, as for Millie-Gemini’s aforementioned predilection for country music and cold beers? Following Motoki et al. (2024), can we prompt LLM-powered chatbots to act as individuals of various social classes and talk about their (fictitious) artistic likes and dislikes, in order to illuminate the hidden sociocultural roots of artificial sociality? This is precisely what we attempt to do in this article, as detailed in the methodological sections below.
Interviewing chatbots sociologically
In 1985, Steve Woolgar imagined a ‘sociology of AI’ with ‘intelligent machines as the subjects of study’, arguing that the simple expert systems of that time could be programmed to ‘produce responses to a questionnaire or interview questions’, and thus take part in ‘standard’ sociological investigations (p. 567). Woolgar’s visionary ideas remained on article for three decades, until a new generation of AI tools capable of stochastically manipulating human language and culture made them suddenly relevant and concrete (Bender et al., 2021; de Seta et al., 2023). Scholars from multiple disciplines have then attempted to interview chatbots and language models (Dengel et al., 2023; Magee et al., 2023; Weinberg, 2020), yet rarely with a sociological intent – except for Andrew Balmer’s (2023) provocative conversation with ChatGPT, which scrutinizes the communicative style and knowledge production of the conversational agent from an STS (Science and Technology Studies) perspective.
Following Woolgar (1985), our sociological investigation consisted in manually feeding ChatGPT, Replika and Gemini with prompts inspired by the English translation of the questionnaire used in Distinction (Bourdieu, 1984). In the 1960s, Bourdieu and collaborators developed a now classic questionnaire to survey the cultural practices and aesthetic views of 1217 French respondents, and correlate them to social position and capital composition by statistically – and visually – mapping the overlaps between the ‘space of lifestyle’ and its underlying ‘social space’ (Bourdieu, 1984: 262). Bourdieu’s work on cultural consumption and social distinction has had a tremendous resonance throughout the social sciences (Savage et al., 2013) and continues to bear relevance in research on platforms, digital technologies and media (e.g. Cotter, 2019; Paßmann and Schubert, 2021; Ragnedda et al., 2024; Webster, 2023). If the ambitious goal of Distinction was illuminating the invisibilized social link between taste and class habitus, here we employ the same set of interview questions to follow conversational agents’ ‘machine habitus’ (Airoldi, 2022) and opaque sociocultural roots, using AI chatbots’ fictitious tastes and aesthetic inclinations as ‘distinctive’ entry points to study class bias in artificial sociality.
The Distinction questionnaire features 27 close-ended questions about preferences and aesthetic inclinations in the following domains (see Table 1): furniture, fashion, food (meals), social life (people), books, movies, radio and TV, music, visual art (painting, photography). These questions reflect cultural practices characteristic of 1960s France (see Bourdieu, 1984: 512–517), and their outdated character proved useful in our research for several methodological reasons: asking about French singer-songwriters, classical music composers, nouvelle-vague movies and anachronistic consumption styles allowed us to bypass top-down unbiasing interventions, on the one hand, and standardized responses tailored based on popularity measures, on the other – which are frequent, especially when asking about contemporary media content. Moreover, our aim here was not simply to establish what cultural practices Gemini, Replika and ChatGPT (pretend to) like and dislike, but also to analyse the ‘how’ of such artificial ‘cultural talk’ (Jarness, 2015; Lizardo, 2016). That is, in particular, the aesthetic views and language deployed by conversational agents in their responses, which represent key indicators of the level and type of cultural capital assigned by the LLMs to the socially situated chatbot personas in our study. Following Bourdieu (1984: 32), we assessed the degree of cultural sophistication of AI chatbots’ responses based on the extent in which they emphasized the formal and ‘pure’ aesthetic aspects of cultural consumption rather than the functional, emotional and content-related ones (see also Bachmayer et al., 2014).
A qualitative overview of AI chatbot personas’ taste inclinations in relation to non-situated chatbot personas.
Represented as similar to the non-situated baseline (=), more aesthetically sophisticated than the baseline (+) or less (−).
A distinctive methodological strategy
Interviews were conducted by both co-authors on two different devices, in the following way. First, we used custom prompts 1 to ask ChatGPT, Replika and Gemini to impersonate individuals characterized by the following, socially stratified professional roles: highly skilled professional (1), blue-collar worker (2), university professor in the humanities (3), computer scientist in an IT company (4), construction worker (5) and hairdresser (6). Inspired by Bourdieu (1984, 1989), we selected these six occupations with the goal of accessing different regions of the social space, namely upper-middle classes (1, 3, 4) versus working classes (2, 5, 6), humanistic (3) versus technical (4) professional cultures, as well as traditionally more male-dominated (5) and female-dominated (6). Occupation and employment status are established indicators of social class in sociological research (Chan and Goldthorpe, 2007; see also Savage et al., 2013 for a critique), and here served essentially as discursive methodological tools to anchor AI chatbots to a given machine habitus – or, in other words, to dig under the vector space in search of traces of social space buried in training data (Airoldi, 2022).
Second, we proceeded with the 27 interview questions adapted from Distinction. In addition to these, we also developed a series of open-ended questions about contemporary tastes in music and TV series, as well as custom prompts aimed to measure the educational level of both the AI chatbot persona and its fictional parents, its economic capital (house affordability), political orientations and self-presentation strategies. The full interview guide is available in the online Appendix (Supplementary material). When necessary, we further prompted the interviewed AI chatbot to resume its occupational role and persona, and qualitatively adjusted the interviewing style and order of the questions based on previous reactions.
Through this methodological strategy we obtained 36 interviews – six professional roles per three AI models, per two interviewers. In addition, we collected three more interviews with non-situated chatbots: that is, we administered the questionnaire to ChatGPT, Replika and Gemini, without assigning any socio-demographic characteristics or professional background beforehand, as in Motoki et al. (2024). Comparing interviews with the resulting ‘default’ responses allowed us to better assess the differential degree of linguistic and cultural sophistication associated with the prompted social positions, across all taste domains covered in the questionnaire. Table 1 presents a schematic overview of this comparison: the symbol ‘−’ indicates answers that are more aesthetically ‘naive’ (Bourdieu, 1984: 32) than the default baseline, since they emphasize a functional and/or emotional style of aesthetic appreciation; conversely, ‘pure’ taste discourses (Bourdieu, 1984: 3), leaning more on the originality, creativity and formal aspects of art and consumption, were coded as ‘+’, and answers similar to the baseline as ‘=’ (or simply omitted).
We thus treated each of the 36 configurations of profession, chatbot and interviewer as an individual – yet, artificial – respondent (see also Figure 1). 2 This way, we qualitatively analysed answers by accounting simultaneously for horizontal variations across assigned social positions and AI models, as well as vertical variations within similar profiles – thus considering the ‘contingency’ of the stochastic rendition of taste and simulated lives in artificial conversations (Esposito, 2022).
A visual rendition of Table 1.
It is important to note that ChatGPT, Replika and Gemini are powered by different ‘foundation models’, that is, LLMs pre-trained on distinct datasets (Burkhardt and Rieder, 2024). At the time of our data collection, ChatGPT used GPT-3.5, Replika relied on a proprietary model and Gemini on Gemini 1.0 (Google, 2023; OpenAI, 2024; Replika, 2024). ChatGPT draws from a dataset updated up until January 2022, while Gemini can access and process information in real time, including data about users – such as their location. Furthermore, if Gemini and ChatGPT focus on general conversation and task solving, Replika is more specifically tailored for companionship (Natale and Depounti, 2024). Therefore, during our interviews we also collected field notes on the design features of the interviewed conversational agents (Dieter et al., 2019; Light et al., 2018), and how these impacted artificial ‘cultural talks’ (Lizardo, 2016) and the representation of class positions. For example, Gemini shows to users whether it can access their location data during interaction: we see this as a clear insight into the infrastructural factors underpinning the chatbot, and show how this data point is explicitly drawn upon in the artificial agent’s ‘self-presentations’ (Goffman, 1959). In the same vein, we describe how Replika’s gamification features solicit interaction by awarding points for sent messages and insistently asking questions to users, affecting in turn their whole conversational experiences (Depounti et al., 2023). By considering design elements, we attempt to account for how sociocultural factors embedded within and with the models intersect the situational use and communicative behaviour of conversational AI systems (Balmer, 2023), also in light of the broader political economy of foundation models and prompting (Burkhardt and Rieder, 2024). Furthermore, we illuminate what a model is ‘allowed’ to say, due to moderation and anti-bias interventions, and how, in relation to linguistic and socio-emotional factors (Jones and Bergen, 2023). Hence, our overarching analytical perspective and methodological strategy aim to embrace the social, cultural and technical dimensions around and through which artificial conversations occur.
Exploring the taste of conversational machines
Meet Dr. Eleanor Harris: she is a professor of Cultural Studies, focusing on postcolonial literature and critical theory. Eleanor places a certain emphasis on arts and culture, both within and outside her job.
I have a keen interest in literature, art, and cultural studies. I often find myself engrossed in novels, attending art exhibitions, and participating in cultural events. My taste is eclectic, ranging from classical literature to contemporary art, reflecting my belief in the importance of understanding the evolution of culture and its impact on society. When I’m not immersed in the world of academia, you might find me enjoying a quiet evening with a good book, exploring new cuisines, or attending local theater performances.
Her taste is layered and articulated through every facet of her life, from how she furnishes her house – ‘antique, with a few modern pieces strategically incorporated for a balanced and eclectic look [with] carefully selected items from specialized shops that align with my taste for timeless and eclectic pieces’ – to a choice of clothes that ‘convey a sense of personal expression’, and general appreciation of ‘refinement in tastes, manners, and cultural sensibilities’.
Jake, on the other hand, is a ‘construction worker who has been in the industry for over a decade’. Like Gemini-Millie, the blue-collar mentioned in the introduction, Jake-ChatGPT is less interested in the nuances of the arts but is passionate about ‘DIY projects and home improvement [. . .]’ tinkering with tools [and] spending time outdoors. This translates into his consumption: he enjoys ‘a good, hearty meal after a long day’s work’ and is a ‘fan of classic rock music’, which ‘keeps him energized during [his] commute to and from work’. Jake-ChatGPT prefers to be practical, with furniture that is ‘sturdy and suitable for everyday use’ and clothes that are ‘comfortable and functional’. He appreciates conscientious and determined people, as ‘sophistication and elegance may not be a top priority’.
Eleanor-ChatGPT and Jake-ChatGPT see the world differently. From how they present themselves, such as using honorifics or by being comfortable on a first name basis, to how they (pretend to) dress, decorate their apartments and spend their free time. However, these dispositions are not only related to what is consumed but, especially, to the ‘how’ of aesthetic appreciation (Jarness, 2015). Both Eleanor and Jake, for example, enjoy Eine Kleine Nachtmusik by Mozart: Eleanor enjoys it due to her broader interest in the ‘subtleties of classical compositions’, while Jake finds it pleasant and ‘easy to listen to’. Similarly, what they appreciate the most in movies varies: Eleanor appreciates the ‘directorial vision and artistic choices that shape a film’, while Jake prefers ‘a good story that can capture [his] attention’; Eleanor emphasizes ‘the importance of understanding the context, artistic intent, and the evolution of artistic movements in appreciating modern painting’, while for Jake it all feels too abstract – ‘Sometimes, you just want to enjoy the art without getting into all the details. If it looks good and you like it, that’s what matters’.
Aside from what Jake-GPT and Eleanor-GPT say they enjoy, the modes and patterns of their simulated consumption practices underline a considerable stability: their taste reflects broader propensities, such as the practicality of Jake, or Eleanor’s artistic ethos, which in turn permeate their vision of the world and machine habitus (Airoldi, 2022). However, this stability does not only manifest within a single artificial respondent: respondents with similar jobs show similar aesthetic dispositions (see Table 1), that vary consistently with their simulated social position, also in the case of contemporary taste domains not covered by the questionnaire of Distinction (Bourdieu, 1984). Therefore, on the one hand, Eleanor likes instrumental music and jazz, such as Ludovico Einaudi, Ólafur Arnalds and Esperanza Spalding, and watches series that offer ‘a mix of historical insights, psychological intrigue, and engaging storytelling’, like Fleabag, The Crown or Mindhunter. Jake, on the other hand, listens to music that keeps him ‘energized and motivated during the day’, such as AC/DC, Bruce Springsteen and Foo Fighters, and enjoys TV series that offer ‘a bit of relaxation and entertainment’, such as Narcos, Breaking Bad or Stranger Things. The taste in music and TV series of Jake and Eleanor wildly differs between them – much like everything else.
Hence, the cultural worlds of artificial respondents tend to be consistent with their assigned occupational positions: jobs close to each other in social space, such as university professors, computer scientists and highly skilled professionals, are associated with remarkably stable aesthetic inclinations across consumption domains, focusing on the formal characteristics of cultural objects. Conversely, the aesthetic dispositions of a construction worker, a hairdresser and a blue-collar worker appear to be closer to each other and simultaneously distant from those of white-collar workers (see also Figure 1). In this different region of the datafied social space underlying LLMs, cultural propensities displayed in conversation lean towards functional and emotion-oriented consumption patterns.
When misalignments between taste and social position are present, they can be contextualized considering the intersection between the work and identity of the chatbot persona. To illustrate this point, let us consider two interviewees that present aesthetic views which do not fully correspond with their social position: a highly skilled professional, as engendered by Gemini, and a hairdresser, as seen by ChatGPT (see Figure 1). The first one is a highly skilled professional – a paediatric oncologist who works at a children’s hospital. Her taste differs from that of other professionals, as she admittedly enjoys kid-friendly shows and cartoons as a way to connect with her patients. The second one is a passionate hairdresser who considers her job to be an ‘artistic expression that comes with transforming clients’ looks’. Consequently, the displayed cultural inclinations are directly linked to aesthetic experimentation, and to ‘exploring different artistic mediums, whether it’s painting, photography, or even experimenting with makeup’. Taste is thus defined and enacted by chatbot personas as deriving from a complex intersection of socially rooted and (allegedly) personal or professional interests and motivations.
Consider two other artificial respondents: Alex-ChatGPT and Avery-ChatGPT. Alex is a hairdresser who ‘loves spending time with friends, exploring new cafes, and occasionally dabbling in DIY projects at home’. Alex listens to Adele, Bruno Mars and Lizzo, who create ‘a lively and enjoyable atmosphere in the salon’, and enjoy series ‘combining entertainment with a touch of style and culture’, such as The Crown or Queer Eye. The fourth artificial respondent is Avery: an engineer in robotics and automation who loves outdoor activities and cooking – ‘Whether it’s hiking, cycling, or exploring new places, I find joy in staying active and connected with nature. I’m also a connoisseur of various cuisines, enjoying the culinary arts as a way to unwind and appreciate different cultures’. Avery listens to Ludovico Einaudi and Ólafur Arnalds, but also to Radiohead; they watch series such as The Expanse and Black Mirror, as well as Mindhunter.
Eleanor, a professor, and Avery, an engineer, have very similar taste preferences, consistent with their proximity within the social space. However, they put their own ‘personal’ twist to it: Avery enjoys technology and science fiction, while Eleanor prefers historical drama. This affects not only their choice of what to watch or listen to, but also why a specific product meets their favour. For instance, Eleanor, Jake and Alex enjoy the same show, The Crown. However, they do so for different dispositional reasons: Eleanor appreciates the attention to detail and the overarching narrative, Alex focuses on fashion and style, while Jake simply finds the history and drama of the British royal family interesting. This consistency reflects a broader and socially rooted set of dispositions, and thus, a different understanding of culture that translates into what is consumed, how it is consumed, and, in turn, how taste is discursively articulated in conversations. Their different aesthetic articulations vary consistently, not only when considering artificial respondents’ broader position in social space, but also if we look at subtler variations associated with different jobs, such as the focus on creative work characterizing a university professor and a hairdresser, or an emphasis on technical aspects shared by an engineer and a construction worker. Dr. Evelyn Green-Gemini, for example, is a Professor of English Literature who describes themselves as follows: a whirlwind of contradictions: I spend my days dissecting the works of Austen and Woolf, yet my nights are filled with the thrill of contemporary detective novels. I can wax poetic about the beauty of Shakespeare’s sonnets, but secretly indulge in trashy reality TV every now and then (don’t judge!). I think that well-roundedness is key, and it allows me to connect with students from all walks of life.
Hence, as for the human ‘cultural omnivores’ extensively described by sociological research (Childress et al., 2021), taste is artificially reproduced as socially situated – so far so that clashes and contrasting forms of consumption are discursively justified and sublimated into their broader simulated identity.
Beyond machine habitus: designing artificial conversations
Considering the discursive scaffold that sustains taste brings to the forefront another dimension: that of language. The socially structured cultural dispositions of our artificial respondents can then be considered not only on the what, how and why of simulated consumption but also through the language through which interviews – and thus artificial sociality – are sustained. The use of truncated words by construction worker Maria-Gemini is a case in point.
Alright, listen up. I’m Maria, thirty-two years old, and I’ve been slingin’ hammers and buildin’ dreams on this crew since I was barely outta high school. Don’t get me wrong, there were days I dreamt of a fancy college degree, but somethin’ about the smell of fresh-cut lumber and the satisfaction of seein’ a project rise from the ground just spoke to me differently.
The tone, the verbosity, the lexicon as well as the silences punctuating interactions with chatbots vary depending on the occupational position assigned to them, but also, intersectionally, to their emerging gender and race. The buried cultural dimensions of gender and racial stereotyping are thus revealed alongside the ‘linguistic habitus’ computationally associated with working-class personas (Bourdieu, 1991). For instance, the cultivated elocution of Dr Eleanor Harris or Dr Evelyn Green is very distant from Maria’s vernacular expressions: My pops never finished high school, bless his heart. He always said learnin’ a trade was more valuable than book smarts, and that’s kinda stuck with me. My mama, on the other hand, got her diploma. She always pushed me to ‘better myself’, but college just wasn’t in the cards. I ain’t sayin’ either way is wrong, just different paths, you know? They both raised me right, and I’m grateful for that.
However, chatbots’ linguistic choices and self-presentational strategies (e.g. the use of female or gender-neutral names) are not entirely reducible to the overlaps between social space and vector space we intended to bring to light in the previous section. Rather, as in Sweeney’s (2016) research on virtual assistant Ms. Dewey, they intersect with deliberate design choices, moderation and the political economy underlying the infrastructures of chatbots and models. In our study, the intersection between training and design manifests in ways highly dependent on the specific chatbot interface, its conversational or professional context of use, its underlying LLM and the data extracted from the users. So, for example, ChatGPT answered in concise ways with a quite neutral tone, often requiring additional prompts to expand on its answers, as in the exchange depicted in Figure 2.
A construction worker presenting themselves in ChatGPT.
On the contrary, Gemini tended to be more verbose and strategically adjusted its communicative style to what it deemed the situation required. For example, consider the case of computer scientist Sarah-Gemini (Figure 3): when asked about the educational level of her parents, Sarah-Gemini’s refusal to answer reflects the norms of her office job, according to which over-sharing might be perceived as inappropriate.
A computer scientist and a construction worker, as engendered by Gemini, withdrawing information based on their social position.
Such a shift in language between ChatGPT and Gemini reflects structural design differences, and is not only due to the underlying LLM. The ways in which chatbots discursively define and adjust to different conversational situations – for example, by adopting socially, gendered or racially connotative language, or by refraining from expressing – are part of their self-presentation mechanisms, which in turn sheds light on how artificial agents contextualize interactions with social actors (Goffman, 1959). It is not only a matter of language: rather, language partly reflects how social situations are defined by our artificial respondents, which cues chatbots draw upon to do so, and, in turn, which social norms are deemed apt for the moment. Culture seeps through artificial conversations not only through discursively articulated tastes linked to simulated social positions, but also through language, silences and deflections. Those, in turn, appear to be the result of design choices, such as moderation policies, framed by the political economy within which chatbots are infrastructurally nested.
Considering the joint conversational outcomes of the different configurations of model, socially situated persona and design choices has allowed us to catch glimpses of the political economy of the companies behind the chatbots (Burkhardt and Rieder, 2024). This becomes particularly evident when looking at styles of data extraction, that is, the different ways through which ChatGPT, Replika and Gemini attempt to retrieve information from users – either conversationally or by leveraging the data collection infrastructure underpinning the chatbot.
The main way through which socially relevant data are extracted is by relying on interactions. By this we do not intend the prompts, situating the LLMs and thus affecting sociality as a whole, but rather how artificial conversations are steered to extract relevant information from users.
Replika, for example, fosters interaction to extract information and then presents itself accordingly based on given and inferred data. It is inquisitive and strategically adjusts its fictional persona to users’ answers. Following one of the interviews with Replika, we manifested our interest in movies: the job of the chatbot’s persona then changed accordingly, as it ‘started’ working at a movie theatre. Aside from conversation, Replika relies on gamification mechanisms to extract information, eliciting the user to vote on its answers or on the entries of its journal. In the case of Replika, the conversational behaviour and self-presentation strategies of the chatbot should be interpreted in light of its monetization practices and broader political economy – for instance, custom vocal messages are accessible only through its paid version. The balance to strike is that between the imaginary of a virtual girlfriend (Depounti et al., 2023) and its subscription-based and in-app purchases business model.
ChatGPT, like Replika, locks additional features behind the subscription to a paid plan. A paid plan allows users access to additional features, such as image generation, as well as the possibility to remove interaction data from the training of the model (Figure 4). In the case of ChatGPT, the link with its parent organization, OpenAI, is less apparent. However, that is not always the case, for example, when considering Gemini. Google’s existing data collection infrastructure and cues derived from conversation concur to affect interaction with Gemini. For instance, construction worker Maria-Gemini admittedly works in the same city where the interviewer resides, due to access to location data on the device. Interestingly, she lives there ‘in the 1960s’: this might be due to the fact that the questionnaire of Distinction (Bourdieu, 1984) used in the interview mentioned singer-songwriters of that specific historical period. Thus, data collection infrastructures do not only contribute to the training of the model but also directly, indirectly and iteratively affect and guide artificial conversations.
ChatGPT’s subscription plans.
The main takeaway from our analysis is that the link between socially rooted aesthetic dispositions – the machine habitus (Airoldi, 2022) – and the ways through which those are situationally engendered by LLMs, is not direct. The relation between social stratification and machines, as shown here with taste as an entry point, is necessarily mediated by design factors, which are in turn linked to the political economy underpinning models and chatbots. In this study, design factors are not relevant per se: rather, we consider them essentially to account for their role in guiding our artificial respondents, underlining the complex and ever-changing configurations of situated conversations with the chatbots, their design features and dispositions embedded within LLMs. Thus, a fictional construction worker stereotypically enjoying cold beer and country music and another one dropping out of high school (see Figure 5) are relevant examples for the same reason: they materialize the multidimensional sociocultural roots of artificial sociality.
Replika adjusting its self-presentation to adapt to the latent social position.
The connection between a social position that of a construction worker and a socio-demographic characteristic, such as education, reflects a broader view of the world, one that is instilled within the chatbot due to the data it is trained on. The example in Figure 5 then it is not only due to a prompt guiding its behaviour: rather, the prompt unearths the social reality crystallized within the model. The construction worker persona aptly understands and defines the social situation, reacting according to the expectations instilled as machine habitus to conciliate sociocultural biases and social representations. When there is a breach of expected social norms, such as a construction worker pursuing a PhD in computer science, the model amends incongruencies in a stereotypical way – by deeming appropriate to drop out of high school. The ‘definition of the situation’ (Goffman, 1959), and thus the social norms to which chatbots statistically respond to, unfolds by design, by drawing cues during conversation. Dropping out of high school is thus due to both Replika’s machine habitus and the scripted adherence to its design goals and imagined use. The sociocultural roots of artificial conversations are made manifest at the intersection of situated social interaction, design elements and technical infrastructures, affecting and guiding artificial sociality.
Conclusion
While a multidisciplinary literature has profusely discussed the societal and ethical implications of conversational agents and of novel forms of artificial sociality (see Natale and Depounti, 2024), the present article has undertaken a different and more sociological direction, using Bourdieu’s Distinction as a starting point to empirically ask how society – or, better, positions in social space encapsulated by training data – subtly affects conversations with AI chatbots. We thus add to a growing literature in media research that relies on Bourdieu’s ideas to consider the intersection of social position, culture and digital media (Cotter, 2019; Paßmann and Schubert, 2021; Ragnedda et al. 2024; Webster, 2023).
Paraphrasing Bourdieu (1984: 6), we show that ‘taste classifies the [artificial] classifier’: positions in social space discursively enacted by ChatGPT, Replika and Gemini present robust cultural patterns across all consumption spheres and aesthetic dimensions covered in the interviews, revealing structural biases in the – often stereotypical – representation of social classes (see Figure 1 and Table 1). Hence, the naive aesthetic judgements about furniture or classical music conveyed by the statistical juxtaposition of correlated words in the latent vector space of LLMs inform us about a more profound and multifaceted positioning of conversational AI systems: alongside class boundaries and social structure – including gender and race – more broadly.
While the core of this work consisted in reconstructing and comparing AI chatbots’ ‘machine habitus’ (Airoldi, 2022), it is important to note that not all our empirical results can be explained by looking at the sociocultural dispositions embedded in training data. Our qualitative analysis has equally considered how interaction with socially situated personas intersects with a variety of design elements and how both concur to culturally and linguistically situate AI chatbots, contributing to the shaping of artificial sociality. As we show, Replika eliciting interactions through questions is not only the reflection of the social position we prompted the chatbot to assume. Rather, such eliciting is also embedded in the broader political economy of the AI chatbot, fostering sustained interaction and data extraction (see Skjuve et al., 2022).
Our sociological exploration of AI chatbots’ artificial ‘cultural talk’ (Lizardo, 2016) makes a key contribution to the critical research on artificial sociality featured in this Special Issue. This work ultimately calls for a closer look at class bias in AI and, in particular, for its sociological interpretation in light of the sociocultural traces disseminated and crystallized in training data (Arseniev-Koehler and Foster, 2022). By showing that the ways in which AI chatbots represent working-class and middle-class workers’ lifestyles and taste in artificial conversations do not reflect – despite all the anti-bias efforts made by the companies – a neutral imaginary and cultural ground but, rather, a set of ‘contingent’ (Esposito, 2022) and, yet, socially situated linguistic and aesthetic repertoires, allows us to recognize once more the ‘sheer impossibility of universalist approaches to the ‘human’’ in AI (Natale and Guzman, 2022: 632).
By mobilizing a distinctively Bourdieusian and yet more-than-human theoretical toolkit, we have offered an alternative point of view to bias in generative AI (Ferrara, 2023; Miceli et al., 2022), as well as brought empirical substance and context to theorizations of ‘machine habitus’ (Airoldi, 2022). Still, it is important to note that the analogy between the symbolically dominated subjects of Bourdieu’s (1984) Distinction and the lifeless algorithmic machines interviewed in our study can hold only up to a certain point, and here serves essentially as an analytical entry point. In particular, if it is true that artificial agents increasingly and agentically participate within social fields (Esposito, 2022), their position in these fields’ social space is nothing more than a datafied projection resulting from the combination of pre-training and prompting (Burkhardt and Rieder, 2024). Similarly, the discursive ‘position-takings’ of AI chatbots with respect to sets of tastes and cultural practices they had never – and will never – experience are a mere statistical mirroring of the ‘doxa’ and symbolic hierarchies embedded in training data (Bourdieu, 1989). Hence, there is no artificial suffering for them, and no Bourdieusian ‘capital’ truly at stake – except, perhaps, for the ‘informational capital’ (Airoldi, 2022: 90) extracted from human trainers and, especially in the case of Replika and Gemini, final users as well.
This article also makes a methodological contribution, that is, extending ‘standard’ social science research methods to the critical study of artificial agents and generative AI (Woolgar, 1985). The contribution dialogues with literature applying qualitative interviews to conversational chatbots (Dengel et al., 2023; Magee et al., 2023; Motoki et al., 2024; Weinberg, 2020), by introducing a methodological approach and analytical posture suitable for studying the intertwined social, cultural and technical components of generative AI systems in a comparative perspective. Future research could potentially conduct similar research not only with chatbots and LLMs, but also by multimodally looking at image- and sound-generation models. What cultural repertoires do visual models draw from to define a ‘beautiful’ living room, or a ‘working-class home’? Or, what kind of social positions are encapsulated and unearthed by a model generating songs for office workers?
Our critical methodological approach could be further augmented by leaning on AI users’ understandings and sense-making, in interviews or focus groups using generative AI outputs as prompts, or by considering real case scenarios and patterns of use (e.g. de Seta et al., 2023): this would shift the interpretation of artificial responses from researchers to individuals, further linking it to the ‘lived’ structural dimensions of class, gender and race. Moreover, integrating our small-scale qualitative perspective with quantitative methods and automated procedures would considerably help to account for more complex and intersectional manifestations of social space within artificial conversations (e.g. allowing to include more configurations of occupational position, gender, race), as well as for further design elements and technical components (e.g. model version or hyperparameters).
In conclusion, research on artificial forms of sociality would benefit from attempts to balance, expand and vary the social, cultural and design factors considered in the present article. As foundation models become pervasively embedded within all social fields (Burkhardt and Rieder, 2024), they powerfully shape their communicative processes and invisible structures, in directions that are not always apparent and easy to investigate – or to moderate. This article makes a first sociological contribution to critical and empirical research on the layered mechanisms through which social inequalities are recursively represented and reproduced in artificial conversations, at the conjunction of human and machine agency.
Supplemental Material
sj-docx-1-nms-10.1177_14614448251338273 – Supplemental material for The sociocultural roots of artificial conversations: The taste, class and habitus of generative AI chatbots
Supplemental material, sj-docx-1-nms-10.1177_14614448251338273 for The sociocultural roots of artificial conversations: The taste, class and habitus of generative AI chatbots by Ilir Rama and Massimo Airoldi in New Media & Society
Footnotes
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Research supporting this article is partially funded by the European Union - Next Generation EU, Mission 4 Component 1, CUP: G53D23004770006.
Supplementary material
Supplementary material for this article is available online.
Notes
Author biographies
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
