AI and narrative revival? The story of working with a Chatbot on the secondary analysis of German-Polish borderland interviews and insights for qualitative sociology on the verge of a socio-technical transformation

Abstract

The paper examines the use of chatbots for secondary analysis of qualitative interview content, focusing on narrative research. It utilises interviews with leaders from communities along the Polish-German border to explore socio-spatial transformations as test material. AI was used to recode data concerning de/rebordering processes, identifying three ways of narrating the border. The quality of the analysis is further tested by predicting how these narratives might evolve due to influences such as a pandemic and war and by exploring the potential for managing them in upcoming cross-border renewable energy projects. The paper highlights the potential and prerequisites for successful AI-supported analysis but also raises questions about the use of artificial intelligence in achieving qualitative interpretative depth, calling for a deeper discussion within the research community on how to integrate AI effectively. Referring to the tradition of sociological research on knowledge and artificial intelligence, the paper also suggests further studies on the anticipated transformation of research practices in qualitative sociology and outlines key areas for conducting such studies.

Introduction

For several years, my colleagues and I have been studying re/debordering processes (Hennig, 2021) in the Polish-German borderland, focusing on the twin cities of Frankfurt-Słubice and Gubin-Guben and the surrounding rural areas (Frąckowiak et al., 2021; Rogowski and Frąckowiak, 2023). Key events coincided with data collection and processing: the rightward shift in Polish politics, the COVID-19 pandemic, the humanitarian crisis on the Polish-Belarusian border, the war in Ukraine, an ecological disaster on the Polish-German border river Oder (Gottwald et al., 2024), a European farmers’ protest, the recent reinstatement of border controls by Germany and Poland. These events enhanced the Polish-German border’s role as a ‘laboratory for European integration’ (Schultz, 2005), stirred the experience of free flow and restored the image of the border as a state-maintained line separating external threats (Kajta and Opiłowska, 2023: 1319).

I revisited transcripts from the in-depth structured individual interviews we conducted in 2018 with local experts to better understand the conditions for recent rebordering. Having worked on the paper which focused on this very issue (Frąckowiak, 2024), I have decided to follow other researchers’ suggestions (Opiłowska et al., 2022: 44) and view re/bordering processes as evolving, interconnected narratives that frame human experiences over time and thus co-shape the transformations of the Polish-German border. Facing challenges in viewing the material once more, but otherwise, I decided to test the capabilities of artificial intelligence for this purpose. The scarcity of data (the sample consisted of 9 expert interviews) imposed certain restrictions but also enabled a thorough understanding of the research material, allowing better evaluation of generative AI’s potential.

The unexpected results of the first attempt encouraged me to delve deeper into the usability of chatbots in qualitative analysis. Given the gap in sociological studies, this goal seemed even more justified: as Rezaev and Tregubova (2023: 7) point out, artificial intelligence is predominantly studied within the frameworks of communication studies, science and technology studies, and critical theory. However, these analyses are rarely translated into methodological reflection concerning specific research projects (Rezaev and Tregubova, 2023: 7). In qualitative research, notable examples explore experimental uses of ChatGPT and compare AI with human coding results (e.g. Kocoń et al., 2023; Morgan, 2023; Zhang et al., 2024).

This paper explores the practical use of large language models in qualitative analysis, contributing to the growing field through a differently designed experiment. I approach ChatGPT not from a data science angle, but as a borderland researcher. The aim is to assess its potential within an iterative framework, where researcher-generated codes are repeatedly processed by the chatbot, not only through descriptive (thematic) coding, but also via interpretation, speculative reasoning, and narrative forecasting. Designed as an open-ended process with no predefined assumptions, each step of the following experiment builds on the previous one. The first part of the paper focuses on the analytical outcomes of using the chatbot to interpret and project borderland narratives. The second part opens with a perspective on a potential revival of narrative analysis in qualitative research triggered by the new possibilities, and then critically assesses the capabilities, limitations, and ethical concerns that goes along. Assuming every new tool invites reflection, I adopt a contemplative, essayistic tone to convey an experience that may resonate with researchers in other fields.

Analysis using a Chatbot: The case of research on the Polish-German borderlands

I approached the experiment equipped with files exported from MAXQDA, containing categorised and anonymized excerpts from interviews we had conducted. Following the axial coding paradigm in grounded theory (Strauss and Corbin, 1990), this included speaker-tagged quotations categorised and annotated with the following codes: causes, context, and consequences of debordering and rebordering processes along the Polish-German border.

The next step was to further analyse this coded material using a customised GPT-4 chatbot (OpenAI). To reduce hallucinations, I used a strategy like that of Zhang et al. (2024). The key was to enrich the model’s context to ensure that prompt engineering, together with the neural network’s attention mechanism, did not compromise analytical quality, as discussed by other scholars (Ji et al., 2022; Kocoń et al., 2023; Morgan, 2023). Prompts first set the analytical context (qualitative interview transcripts), then requested a semantic analysis of the border using a prepared set of selected quotations. The prompt instructed that the output be preceded by a review of the categorised excerpt files and include references to their original interview sources.

I applied a few-shot learning approach, guiding the model from simple to complex tasks instead of requesting a full analysis at once. After completing the semantic analysis of the border, I prompted it to reconstruct key narratives about the Polish-German border (Step I). I then extended the session to forecast narrative dynamics in crisis contexts (Step II), and finally, to speculate and model a narrative around a hypothetical cross-border investment (Step III). I provide further details on prompt wording and bot configuration in the referenced sections.

Given the opacity of large language models and the non-deterministic nature of neural networks, full transparency, replicability, and output validation remain difficult (Rossi et al., 2024). Due to the specifics of the analysis presented below, I did not compare the model’s coding with my own or conduct inter-coder reliability tests. Instead, I ensured analytical quality by following best practices from the literature (Ji et al., 2022; Morgan, 2023; Zhang et al., 2024) thorough knowledge of the data set, avoiding analysis types prone to hallucination (e.g. summarising large unstructured data), using categorised quote files, verifying cited excerpts, and empirically cross-checking AI-supported results with external sources. When this was not feasible, I relied on the ‘believability’ of the outcome (Rossi et al., 2024).

Step I: reconstructing narratives of the Polish-German border from interview excerpts

As previously mentioned, I began with initial, simpler analyses conducted on the appropriately prepared interview excerpts. Next, inspired by Czarniawska (2004: 1–16), I prompted the model to identify and describe the properties of the narratives using the categories such as: the topics, basic structure, narrator’s perspective, and potential significance of the processes of re/debordering.

The first narrative developed by the chatbot describes the Polish-German border as a symbol of European integration and cooperation. It emphasises the gradual fading of physical and mental borders between EU nations, towards an ideal of a transnational community. The model characterises this narrative as a story of progress marked by optimism and a pragmatic exchange approach aimed at shared goals. This description blends international and local perspectives, emphasising the fluidity of borders to foster socioeconomic development, mainly through integrating twin cities economically, infrastructurally, and culturally. It turned out that individuals who resonate with this narrative actively engage in inter-institutional cooperation, benefit from cross-border shopping, and pursue transboundary professional or educational opportunities. They also advocate for openness and kindness towards their neighbours across the border. To explain their positions, these individuals highlight their interest in history, language proficiency, and involvement in cross-border projects.

The second narrative about the Polish-German border centres on dialogue and reconciliation. This frame, as reconstructed by the model based on the files with my initial coding and prompts described above, highlights critical themes like the trauma of World War II, redrawing of borders, population relocations, or attitudes towards German buildings now inhabited by Poles and the differences in post-war building cultures. The narrative of reconciliation is characterised by reflexivity, sparked by discussions on the region’s shared cultural and historical heritage, and emphasise how historically burdened cross-border relations can evolve into a shared transborder identity. According to the chatbot, the interviewees who discussed the border in this manner often cite personal experiences, such as growing up in border areas or having familial or collegial ties across the border, which foster a deeper historical and cultural awareness.

The third type of narrative identified and characterised by the model based on data provided portrays the border as a line of division and protection, a guarantor of sovereignty and mechanism for managing illegal migration or cross-border crime. The model suggests that the narrative of defence and protection is subordinate to identifying potential threats linked to national and international conditions. This perspective on the border was the least frequent in our collected material, mentioned only by a professional involved in border surveillance. This is understandable as we conducted interviews during a period of open borders, while the model indicates that narratives of division and protection could gain popularity during crises as awareness of threats and political tensions increases, leading to intensified discussions about the need for security and enhanced control over border movements.

Step II: forecasting narrative dynamics in reaction to crises

Assessing the validity of the above analysis, I can state that the three reconstructed narratives quite well explain the complexity of attitudes towards the transformations of the Polish-German border. For example, Elżbieta Opiłowska’s team conducted research on a similar topic in the exact locations, and also identified the local dominance of narratives about the city with the border and twin cities (Opiłowska et al., 2022: 141–142). These could be seen as variants of ‘my’ narratives on integration and cooperation. In the analysis by these researchers, narratives about reconciliation and protection occupy little space, likely because another narrative dominated the interviews, as in my case.

It would be more precise to say: had explained as I reconstructed the narratives based on data collected before events that intensified rebordering. However, the time that has passed provides an opportunity to observe how these have evolved. I decided to verify this potential by providing successive prompts to the model, following the general guidelines described earlier. To do so, I enriched the model’s context by uploading a file detailing the three ways of narrating the Polish-German border to its knowledge base, and prompted the model: try to deduce how these narratives might change during an international epidemic of a potentially lethal, unidentified flu virus spreading through cross-border mobility.

In such a crisis, the model suggests that the narrative of the border as a symbol of European integration and cooperation would emphasise openness and the need for joint public health and safety actions. The model continues: the European Union is presented as a platform for developing cross-border solutions. According to the model, border crossing restrictions could be viewed within this narrative as a temporary precaution to protect the population. The narrative about dialogue and reconciliation might deepen, seeing the epidemic as another chapter in shared history and an opportunity for cross-border solidarity and joint crisis management efforts, such as health education initiatives and mutual assistance. Meanwhile, the border narrative as a line of division and protection might gain traction as it boosts the need to protect against external threats and justify temporary border closures or resumed controls. However, the model speculates that this narrative could ultimately evolve to stress that health protection requires collaboration beyond borders, not necessarily closing off from neighbours.

Encouraged by the results, I conducted one more trial. This time, I used the following prompt: Based on the analyses so far, let’s speculate further and imagine that border controls are reinstated on the Polish-German border. There is a war in Eastern Europe, and the Eastern German states are concerned about illegal migration. How would the examined opinions and narratives about the border evolve under these new conditions?

The model predicted that in the people’s eyes, the border would transform from a demarcation line into a protective barrier ensuring internal security. It speculated that the integration narrative would be overtaken by narratives stressing the need for protection against illegal migration and threats. This shift in dominant narratives would be reinforced by public discourse, with the media depicting the border as a critical component of security strategy. Tensions and divisions would rise, not only internationally but also within local communities. The model also forecasts emotional disputes between advocates of protective measures and proponents of a borderless Europe. Consequently, quoting the analysis conclusion: ‘Local and transborder identity is redefined, and new challenges test and prompt reflection on the meaning of European solidarity in the face of crises’. Further quoting the model’s predictions: ‘previously flourishing exchange and dialogue are constrained by formalities and physical barriers, which limit the daily functioning of twin cities and, in a longer perspective, lead to the evolution of new forms of cooperation and narrating about the border’.

Given how events unfolded on the Polish-German border during COVID-19, the national narrative of division and protection initially overshadowed the narratives of integration and reconciliation during the pandemic. This dominance was primarily at the national level, likely fuelled by media-induced panic, tensions in Polish-German intergovernmental relations, and efforts of Polish leaders to demonstrate control. Chatbot-driven speculation can again be also validated against existing literature. Studies from the time confirm that the situation was met with significant concern among leaders and local organisations, who narrativised the border closure as a threat to integrated Europe or a tragedy for the functioning of twin cities. The stories of health and safety as supreme values or the primacy of national interests were decidedly less present locally, to use the terms of researchers at that time (Kajta and Opiłowska, 2023: 1325–1331).

As for the more current situation at the Polish-German borders, the reinstatement of German controls did not significantly increase tensions when I conducted these analyses (summer 2023). However, over the past several months, the situation has changed: tightened controls and push-backs introduced after elections in Germany have led to protests and reciprocal measures on the Polish side of the border.

Step III: speculation and modelling the narrative around a hypothetical transborder investment

Considering the forecasting outcome, it is worth noting that I configured the customised GPT model to rely on the research materials. We cannot be certain that the model did not draw on external data (GPT-4 was trained on data up to 2021, including the pandemic period); it is likely that in this case, it primarily relied on the data provided. This is because appropriate contextualization through prompt design – which I aimed to achieve – can steer the model’s probability distributions, guiding the transformer algorithm to generate outputs based on the most immediately accessible content (i.e. the provided research data), due to the relative downweighing of training data and the increased salience of information within the context window.

The predictions mentioned above were also reasonably expected; the AI suggested the most likely outcome as a logical consequence of the materials I provided. Still, it was not without some inaccuracies. Chatbot itself noted that based on the provided research material, the conclusions did not account for external factors like epidemic dynamics, political decisions, or social risk awareness, which proved crucial.

Therefore, AI-assisted analysis of border narrative dynamics might be better suited for managing discourse direction and its impacts rather than precise prediction. This approach aligns with scenario planning, such as qualitative foresight studies (Loveridge, 2008). To explore this possibility further, I conducted an additional analysis using the described model and instructed the chatbot to speculate on cross-border investment in the Polish-German renewable energy network, a topic that intrigued me for several reasons: such a network is indeed planned in the twin city of Görlitz-Zgorzelec (Caron, 2024; Wasielewski, 2024), exemplifying local leaders’ efforts to strengthen cooperation and blur the Polish-German border despite an adverse atmosphere. In addition, it relates to the ongoing energy crisis and Polish-German-Czech tensions around the environmental impacts at the nearby Polish coal mine located in the borderlands (in Turów). Studies on these issues have identified interesting discursive strategies, such as scaling, networking, and territorialisation, used by various narrators to enhance their position in debates over job security, a favourable future, or environmental quality (Bembnista and Gailing, 2024).

The bot was configured as previously. The prompt was: how could those three reconstructed narratives frame joint cross-border energy investments? Also, how they can slow them down, and how to mitigate those risks. In other words, I was interested not only in the analysis of the challenges and benefits but also in identifying likely fields of ‘narrative battles’ (Opiłowska et al., 2022: 49).

Based on the data provided and within the context created by previous exchanges during the same session, the model concluded that the narrative of the border as a symbol of European integration and cooperation would expand with such an investment to include a new topic: joint renewable energy projects and the construction of energy infrastructure supporting green transformation. The model also states that these investments would be narrativising as exemplifying cross-border cooperation for the region’s climate goals and energy security, symbolising the benefits of Polish-German cooperation in line with EU ideals. The model also predicted threats to the narrative of European integration and collaboration. Paraphrasing its findings could include benefits unevenly distributed between countries and investments executed by large corporations at the expense of local entrepreneurs and communities. To minimise this risk, artificial intelligence suggests that joint planning and implementation of investments could enhance narrative cohesion.

Within the narrative of reconciliation, the chatbot predicts that such investments would primarily be discussed in terms of the region’s industrial heritage, viewing them as opportunities to overcome historical resource conflicts through cooperation towards sustainable development. The model also identifies risks, suggesting that this narrative could reignite historical antagonisms and associate the investment with threats to the sovereignty of both countries, appearing as an attempt at domination by one side over the other. To mitigate these threats, the model suggests acknowledging these historical issues and incorporating dialogue with local communities with such concerns into the investment planning process.

The third narrative, as reconstructed by the model, viewing the border as a line of division and protection, would interpret the cross-border energy investment as an opportunity to secure energy supplies, enhance energy independence, and protect against outside EU external energy disruptions. The model suggests that such investments would primarily be seen through the lens of national security. However, it identifies risks such as threats to the natural environment, using investments as tools of political influence, and potential conflicts arising from divergent national energy interests. To counter these threats, the model recommends making joint agreements and regulations on environmental protection and safety transparent, which could strengthen trust between partners and local communities regarding the investment.

Discussing hopes and fears towards AI-assisted qualitative sociological research

I consider the conducted experiment a moderate success. It validated the approach of viewing the transformation of the border as a collection of evolving narratives, thereby helping to advance the processual theory of border transformations, which the de/rebordering dichotomy had somewhat confined. As we know from relating the findings to other studies, the model proved effective in reconstructing these narratives – they help explain the complex relationship to socio-spatial transformations in the Polish-German borderlands and, in some cases, even turned out to be prophetic. It is worth noting that the AI-reconstructed narratives reflect not only the frameworks within which residents think about the border but also align with the primary current border studies approaches, which Opiłowska’s team describes as focusing either on ‘historical heritage and memory studies, the process of European integration, and everyday studies at the borders, [or] the mutual perceptions of border residents’ (Opiłowska et al., 2022: 20–21).

Of course, I could further compare the quality of analysis with and without AI to underline its effectiveness, but such a comparison might reinforce an unnecessary separation between AI-driven and traditional research methods instead of considering how the former may intertwine with the latter. As a qualitative researcher, I was more interested in establishing a personal stance on artificial intelligence’s technological revolution in our research field. As stated, I have decided to share these experiences more broadly because narrative methods will probably regain traction in qualitative sociology due to the contemporary social anxieties and possibilities of Generative AI, but this potential will require a joint effort to rethink the qualitative research practice along the way.

The AI-driven narrative Re-Turn in qualitative research?

The title may seem provocative as large-scale data analysis, especially with computational tools, remains popular since some time. Techniques like topic modelling, sentiment analysis, linguistic inquiry and word count have been widely used to identify patterns in areas such as local food policy (Mazzocchi et al., 2023), energy policy (Debnath et al., 2020), or menopause representations (Rowson et al., 2023). In this sense, transformer-based language models may only boost interest in such an analysis by addressing the long-standing challenge of ‘capturing distant relationships’ between words that shape meaning (Kocoń et al., 2023).

However, raising the title question I was thinking more of qualitative research and the ‘narrative turn’ as popularised, for example, by Czarniawska. Twenty years ago or so, she pointed out that narratives found in everyday conversations or media are crucial as they serve a fundamental role in human communication (2004: 3), allowing to organise people’s experiences into a cohesive biographical whole, but also share them with others to negotiate the common understanding in the absent of the Enlightenment metanarratives (12-14). Due to its limitations in explaining the evolution of social arrangements, the narrative approach has recently receded somewhat in favour of more relational practices more suited to grasp the collective role of embodiment or materiality (e.g. Shove et al., 2012). However, concerning the transformations of values and the prevailing uncertainty in recent years that have spurred individual and social reflection (Marody, 2021: 39-43), we may predict the big narrative ‘re-turn’ after the social urge to weave new stories as a sense-making tool. For example, the importance of discourse analyses in understanding the liminality of borders and narratives during the pandemic has already been acted on by scholars such as Bembnista (2024: 113–114) or Kajta and Opiłowska (2023: 1323).

A second reason for a potential renaissance of narrative analyses in qualitative research is the development of artificial intelligence, particularly attention-layer-based architectures and intuitive, dialogue-driven interfaces that can support more nuanced text analysis and better align with qualitative inquiry. Aside from rare examples, current research predominantly views AI in qualitative research as analogous to a research assistant, aiding in literature review, thematic coding, and language corrections (Kooli, 2023). Support in transcribing recordings or summarising the transcripts could also be mentioned here. However, as Messeri and Crockett (2024: 58) highlight in Nature, AI is increasingly serving not just as a ‘substitute’ but also as an ‘oracle’ identifying promising research areas, an ‘analyst’ discerning patterns in large data sets, and an ‘arbitrator’ verifying the correctness of analyses.

Ongoing debates exist about AI’s efficacy in these roles and its tendency to make errors or hallucinate. Anyone wanting to use ChatGPT in qualitative analyses must be aware, for instance, of the risk of making up or at least paraphrasing statements reported as coming from research participants (which can be mitigated by applying, for instance, the measures I described in the introduction to the previous section). There are also concerns regarding algorithm biases and the ethical considerations it raises regarding the disclosure and description of its use (Christou, 2023b; Van Noorden and Webb, 2023). Recalling concerns about misuse, such as plagiarism or privacy violations (Nah et al., 2023: 284). Worries also revolve around deskilling researchers by dulling cognitive skills (critical thinking, assessment, and judgement) and lowering the number of jobs for those just entering the academic field (Christou, 2023a; Kooli, 2023). Further, foresight studies involving ERC grantees have raised additional worries about AI’s lack of transparency and replicability, loss of creativity and diversity among the academics, and the risk of hyper-productivity due to the acceleration of once time-consuming task completion it enables (ERC, 2023:; Prillaman, 2024: 9–11).

Analysing the line of criticisms, the reluctance towards AI in narrative studies or qualitative research in general may be even more significant. Maybe not so much in AI usage, but in disclosing the fact (Christou, 2023b) and integrating technology into research practices more effectively. But instead of seeing AI merely as a replacement, a scientific fraud, or a threat, I am more committed to exploring how this tool could humanise us by revealing the expectations we place upon ourselves (Krajewski, 2008). In other words, let’s discuss the fears of subjectification and the possibility of enhancing the qualitative and human aspects valued in this field. Inspired by Steinar Kvale (1996), who addressed common critiques of qualitative research in 1996, I would thus like to draw on my experiences below by addressing three deceptive assumptions that have influenced my initial aversion to not employing AI in qualitative analyses. These involve the researcher’s role, the quality of analysis when AI is used and the ethical concerns it highlights.

Premise I: AI will replace qualitative researchers

This concern echoed earlier worries about statisticians during the rise of software like SPSS and, more recently, with plugins like GitHub Copilot, which suggest code for the popular R computing environment. However, while such technology shifted the focus of competences, it did not eliminate the need for statistical expertise. I am expecting a similar shift within the qualitative analysis of interview materials.

With AI-assisted capabilities for preliminary data processing, the recognition of labour-intensive qualitative research tasks may decrease. However, this doesn’t necessarily mean declining demand for qualitative sociologist competencies. Of course, provided – as emphasised by authors who point out this risk (Morgan, 2023: 9) – that the profession of qualitative analyst is seen as more than just descriptive coding, and that the role of artificial intelligence is understood as extending beyond merely accelerating the task through its autonomation.

As quality becomes more crucial than merely generating a summary of conversations, the emphasis will likely shift towards the skills of eliciting the qualitative data that will become synonymous with the endeavour’s success. This will create opportunities to highlight the value of competencies in deepening responses, interpreting them, and confronting these explanations in dialogue with interviewees. The acquisition by large language models of the ability to propose simple thematic codes for material may also amplify the need for transitioning from descriptive to theoretical coding (emphasising grounded theory building through exploring relationships between codes). The new tools and the indicated shift in emphasis will thus align with the need for researchers to engage in nuanced explanations and the ability to situate findings within the context of a discipline’s body of knowledge rather than merely reporting what research participants have to say on any given topic. This is yet another argument countering fears of the disappearing researcher. If we aim to verify AI’s findings and use it to build more complex coding systems, this inevitably requires a profound understanding of the research material.

All this does not necessarily mean obstacles for junior researchers entering the field of professional inquiry. A reduced demand for low-skilled work may, in fact, be accompanied by an opportunity for faster professionalisation, as iterative exchanges with chatbots and emerging frameworks for such processes can facilitate a more thorough understanding of the principles of qualitative analysis (Zhang et al., 2024: 17).

As I experienced firsthand, working with a chatbot also helps to recognise the importance of the social studies of science and technology notion of the importance of shifting between forms of representation in cognitive activities (Latour and Woolgar, 1986). As described above, identifying connections between bordering processes on the Polish-German border became more manageable when I could quickly transform sets of quotes into a list of codes, the list into a table, and the table into a graph depicting relationships between the codes. Crucially, these transformations facilitated recognising previously overlooked connections between dimensions of the studied phenomenon and visualised entirely random combinations that helped break free from habitual thinking patterns.

The AI-supported analysis could also free up some time for activities focused on popularising knowledge. Thus, the development of chatbot use in the study could paradoxically also bolster public sociology, as advocated years ago by Michael Burawoy (2005), by valuing the sociological elicitation of knowledge as much as the ability to use it, highlighting the importance of this specific way of storytelling for social benefit and self-awareness.

Premise II: AI will simplify the depth of analyses, rendering them superficial

This challenge encompasses several concerns: First, there may be a worry that AI will fragment qualitative research’s coherent and holistic nature of storytelling. That is, it will separate the content of the statements and the experiences they refer to from their causes and consequences, as well as reflection on the significance of the interviewee’s perspective, the possibilities of representing these experiences in language, the performative significance of this narrative for the interviewee’s biography, and so on. However, as I tried to demonstrate, the depth and coherence of analysis depend significantly on the coding strategy and familiarity with the discipline’s body of work in this matter.

Second, I was also concerned that the mere organisation of qualitative research material could replace solid argumentation rooted in it. My attempts suggest that ChatGPT, as a research assistant, tends to ‘take it easy’, striving to satisfy the user by delivering linguistically polished answers that combine information from various sources, creating a persuasive impression of coherence. However, this technology, at least for now, seems to lack self-criticism or doubt regarding its conclusions, a framework focused on seeking Popperian black swans that could instead question than solidify the argument (Popper, 1959), such as excerpts from interviewees’ statements that do not fit, challenge the validity or demand nuance in argumentation. Yet, reflecting on the outcomes of AI-supported analysis, we might also question whether this limitation is unique to the machine or also applies to studies conducted by human researchers. Interestingly, initial suspicion towards analyses proposed by artificial intelligence (avoiding the over-reliance warned against by researchers of this technology, see, e.g. Nah et al., 2023: 284) could foster a broader critical scrutiny and uncertainty of reasoning procedures in general.

The concern could be that AI prioritises standards over reflectivity, which is crucial in qualitative and narrative analyses. This risk pertains to analyses relying solely on AI-generated conclusions, which, if they are not to hallucinate, report rather than interpret what is being said during the interviews. AI may popularise structured individual interviews for ease of processing (AI-powered coding assistance performs better with material organised by topics and responses that stay on point). However, it could also lead to a more random selection of quotes, potentially reducing excessive anecdotalization. While conducting the trial described above, I caught myself selecting different, more vivid quotes and tending to shorten them by cutting parts rather than paraphrasing (as ChatGPT typically does), which fundamentally does not necessarily demonstrate reflexivity if we accept that the role of research is not only to seek what is exciting and striking but also to represent the ordinariness of human experiences.

Also notable could be the fear that replicating what ChatGPT has ‘learned’ in the training stage will replace understanding during the analysis. This can be mitigated, as I experienced, by not using AI to work independently on the raw transcripts but by preparing materials carefully and engaging in multiple exchanges with the chatbot, ensuring analyses are ‘filtered through oneself’ in line with hermeneutic principles. This involves explaining and interpreting answers provided and based on the research material, using our experiences and knowledge, and then relating them to the context. The more we know about the subject of research, the better we know and prepare the material, the more qualitative these kinds of hermeneutic exchanges with the chatbot are, and the analysis remains ‘guided by the researcher’s understanding of the research topic’ (Christou, 2023a: 1796). One might even consider relating these exchanges to a concept of deepening (well-known in qualitative research) by reflecting on how enriching the context through successive prompts provided by the researcher and responses generated by ChatGPT resembles the principle of formulating follow-up questions in conversations with interviewees. Or an ideation process typical of the broader sociological practice, where ‘sociological insights often emerge precisely through critique and dialogue’ (Hau, 2025: 64).

Fifthly, AI-supported analyses like presented above raise questions about how to assess their quality using criteria such as replicability, transparency, and validation. While existing literature offers some guidance, it also suggests that we may have to rethink those very categories (Rossi et al., 2024: 156). For example, asking the customised bot to speculate on joint energy investments in the borderland led to uncertain territory. The output, due its nature, lacks direct empirical confirmation, be it in research material provided, literature, course of events or even publicly available data used to train the large language model, increasing the risk of so-called extrinsic hallucinations (Ji et al., 2022: 4). Nevertheless, this area could be still very fruitful theoretically.

Sixthly, and in connection with the above, concerns about integrity versus serendipity in analysis arise. Contrary to fears, AI can foster inventiveness by facilitating the search for connections and encouraging the formulation of new research questions or hypotheses during the analysis process, often seen as the essence of scientific work (Kooli, 2023). The ability to quickly change representations, such as placing conclusions in a table described earlier, is also helpful in those explorations. The ChatGPT could also act as a ‘distorting mirror’ and de facto makes it easier for the researcher to converse with themselves, but also challenging the researcher to think differently, thus enriching both the analytical process and the results by offering alternative languages or perspectives that can cast the matter being examined in a new light.

Premise III: The increased use of artificial intelligence in qualitative research may promote an unethical and exploitative mode of practising this discipline

Conclusions on this issue will largely depend on our belief in AI’s capabilities and our optimism about human nature, as Rezaev and Tregubova (2023: 6) rightly pointed out. However, I would like to highlight six areas where the development of AI in qualitative research will require rethinking current practices in this field.

The first issue is informed consent and ensuring the privacy of individuals whose often personal experiences we wish to process in the cloud and on servers over which we have no control. Transcriptions should be anonymised beforehand. Platforms like ChatGPT offer the option to turn off consent for training models based on the data we provide, meaning that uploaded files and exchanges with the chatbot are stored only for 30 days to address potential abuse claims. Even if someone does not disable this option, to my knowledge, GPT-4 retains linguistic patterns rather than the entire ‘content’ that would allow the statement to be reconstructed. Thus, if we prepare an appropriate consent form, anonymise the material, and configure our model, the situation appears straightforward. Nonetheless, since LLM architectures are rarely transparent, both non-local and local models (i.e. those other than Retrieval-Augmented Generation, or RAG) shouldn’t be considered safe for sensitive data, as both setups may be prone to leaks (Morgan, 2023: 8). Naturally, developing AI-supported analyses, like software for transcribing recordings or proofreading translated papers, further sharpens our awareness of our ethical obligations towards research participants and how we understand their well-being.

The second dilemma involves copyright, and this time concerns researchers, whose papers may be used (often without proper attribution) to train models, potentially leading to unintentional copyright violations or so-called high-tech plagiarism (see Nah et al., 2023: 286). Legislative initiatives are underway, and new platforms and services are emerging to guard against such risks. However, this latest iteration of the ‘death of the author’, described years ago by Roland Barthes (1977), can be met with mixed feelings. Initially, I was excited by the idea of my works feeding a vast knowledge network – less like books on shelves than a mycelium helping someone to guide through everyday life. This perspective is likely more appealing to those who view the citation-count-based system of academic recognition and science behind the paywall with scepticism. On the other hand, it is hard to maintain a purely romantic view of this phenomenon, especially in the context of works on surveillance or platform capitalism (Rezaev and Tregubova, 2023:; Srnicek, 2016; Zuboff, 2019: 7). But there are more issues from the ISA Code of Ethics still unresolved: how can authors’ rights to avoid decontextualization or correct errors be upheld once their work has been used as large language models training data?

Third, as mentioned above, bias in large language models has sparked considerable debate. Western cultural, racial, gender or sexual biases may stem from training data and surface during qualitative analysis (Kocoń et al., 2023). Approaches to bias vary: model developers aim to detect and mitigate it (Morgan, 2023: 7); some scholars consider it as a feature to explore interpretive frames in studied groups (Rossi et al., 2024: 154); others, see it as a trigger to reflect on broader issues of bias in social analysis (Hau, 2025: 60).

Fourth, working with a chatbot (as I know from personal experience, having doubts about how to mark above the parts fuelled by the model speculation on the borderlands narrative’s possible development) raises questions about the essence of the creative component in our academic writing, which would benefit from broader discussion. We feel, and it is even embedded in submission guidelines, that we cannot directly copy and paste fragments generated out of the blue by ChatGPT. But what if these are analyses based on material we encoded and prompted? Does the situation change if the chatbot, as in the examples I referenced above, starts speculating or referring to its ‘own’ knowledge? Should we then cite or paraphrase? What if we use it, as I have in this paper, to support us with translation, language editing and syntax improvement (based on fragments I authored and under my supervision)? Are these questions, in essence, so different from those we might ask ourselves when writing a report based on interviews and relying heavily on paraphrasing what research participants said during the interviews?

The fourth issue concerns the impact of technologies like the one I used in the analysis above on equality of opportunity in global academia. These technologies are tools that facilitate translations of texts from national languages into English, making it easier for researchers with fewer resources to participate in the global knowledge exchange. However, this digital divide also encompasses access, prompt engineering experience, and accepting such technologies in different academic environments. It is easy to imagine scenarios where consumer-accessible technologies are framed as scientific misconduct in ‘peripheral’ countries, while in wealthier academic centres, alternatives and usage cultures are developed that fully pursue these new possibilities.

Finally, energy consumption is becoming an increasingly crucial ethical issue, especially in research on narratives or qualitative sociological studies. Estimates vary, but all point to the growing use of electricity and water associated with the popularisation of artificial intelligence. Apart from air travel for academic conferences, the environmental footprint has not been a significant concern in the social sciences until now. However, this may change with algorithmically supported analyses, massive text corpora, audio transcription, and text translation. I can imagine a scenario where ethics committees evaluate proposals involving the use of Large Language Models, weighing their anticipated carbon footprint against the significance of the research question.

Conclusion

My experiment leads me to question the idea of chatbots as autonomous coding assistants valued mainly for saving time on a labour-intensive task. Instead, I support what Mark F. Hau calls in his project of augmented sociology ‘dialogic ideation’, a process of interaction between the researcher and the language model (2025). Rather than relying on or personifying chatbots, it is more productive, as Hau suggests, to treat them as agents that, through an iterative process, can support human reasoning and writing. Only then, as this analysis shows, can AI contribute to qualitative research without undermining its depth and reflexivity.

To assess the interpretative richness AI can bring within the qualitative analysis, it’s also crucial to identify which tasks within this iterative, dialogic process benefit from its use. Hau’s (2025) observation proves accurate also during the study described above: ‘the strength of LLMs appears to be breadth, not depth’ (p.65). Within the ideation framework, chatbots work well for descriptive coding, linking codes, providing dialogical support, and even speculative brainstorming that may lead to theory-building. They could also help turn messy researcher thoughts noted during the coding or analysis phase into clearer, developable ideas. However, the human researcher remains essential for initial material familiarisation, summarization (Ji et al., 2022: 17), detailed coding, and organising material into reliable analytical wholes (Morgan, 2023: 9), all while accounting for context (Morgan, 2023: 8; Kocoń et al., 2023).

Considering the experiences described in this paper, the perspective we will frame these opportunities and challenges is also vital. Leveraging the insights from social studies of science and technology seems particularly important. For example, it would be a mistake to attempt to separate conceptual work from technical work again, to rationalise tools like automatic transcriptions, Scopus AI facilitating preliminary searches of literature related to a given query, or new capabilities of virtual assistance in supported coding, summarising, and exploring transcripts in the qualitative data analysis software MAXQDA while trying to preserve our research identities intact.

Drawing on the experiences of Latour and Woolgar (1986), Karin Knorr-Cetina (1999), and others who explored the relationships between scientific cultures and evolving research instrumentation allows us to move beyond equating science with the representation of the world. This enables us to ask not whether artificial intelligence can accurately represent, for instance, the narrative structures used by the communities we study, but rather how this new tool is recontextualised within this specific research environment. Furthermore, as I have tried to demonstrate in the paper, it prompts inquiry into how AI compel the transformation of existing practices, mental models, identities, the understanding of science, its social obligations, and so forth. As described by scholars in social studies of human-media communication, this is often referred to as an ‘ontological provocation’ (Goot and Etzrodt, 2023), a concept also picked up by those experimenting with AI in qualitative analysis. The point is not to see AI merely as a tool, but – openly – as something that can transform our ‘epistemological outlook’ (Zhang et al., 2024: 19).

From the sociology of knowledge perspective, two research paths on AI in qualitative studies seem to me worth exploring. The local implementations of this technology, along with the accompanying socio-technical imaginaries (see, e.g. Bareis and Katzenbach, 2022) that frame what is feasible and ill-advised while using it. I also find further experiments on using chatbots in speculative reasoning and their role in forecasting narrative advancement to be of interest, particularly methods for controlling hallucinations in this process and, potentially, facilitating theory-building as an outcome.