Thinking the unthinkable: Generative AI for collaborative,intra-active policy analysis

Abstract

In this paper, we report an exploratory collaboration in data analysis between a human researcher and Generative AI (GenAI). Specifically, we reveal how human-AI collaboration might contribute to alternative interpretations in theory-informed policy analysis. We draw on Karen Barad’s philosophy of agential realism as an anchor point to consider the ethico-onto-epistem-ological affordances and limitations of incorporating GenAI in social sciences research to guide the human-AI collaborative analysis. The exploratory collaboration reported is based on a case involving how a human researcher (the first author) engaged with Copilot, drawing upon a relevant analytical method, to analyse a Chinese education policy document. Through this example, we illustrate how a more intra-active human-AI collaboration could enable the human researcher to think the unthinkable, that is to think beyond the discursive or conceptual boundaries of his own understanding of the policy logics through the analytical framework. We argue that we can benefit from the analytical conversational collaboration process between AI and human intelligence, rather than simply the AI-generated output per se. This study is significant as it demonstrates a much more ethical, interdependent, and indeed intimate form of human-AI collaboration, in which the agency between humans and GenAI is recognised and respected. This more intimate form of human engagement with GenAI might also yield new ways of knowing and being as researchers in conducting qualitative policy analysis.

Introduction

The advent of Generative AI (GenAI), such as OpenAI’s bot Chat Generative Pre-Trained Transformer (ChatGPT) and Microsoft’s Copilot, has caught the attention of educators and researchers in various disciplinary areas in education. A GenAI bot can be understood as an intelligent dialogue system that is capable of Natural Language Processing (NLP) to generate responses to questions prompted by human users (Zheng and Stewart, 2024). For human users, the process of eliciting desired outcomes from GenAI is often described as prompt engineering. Through such prompt engineering techniques, human researchers can work with GenAI to improve not only ‘efficiency’ but also the calibre of academic research more broadly.

In this article, we explore the affordances and possibilities that attend engagement with GenAI, and how such human-AI collaboration can enhance analytical capacities in the area of education policy studies. We draw upon a case in which the first author, Danson, analysed a specific Chinese policy (Ministry of Education, 2014) in relation to the initiation of national competency-based education reforms within the whole Chinese education system (primary to tertiary education) to exemplify such engagement. However, in contrast with more traditional, rationalistic accounts of human-technology interaction, we refer to post-qualitative ideas as a generative way to consider more active human-AI intra-relations which diffuse the boundaries between those ‘using’ such technologies and the actual technologies themselves. Drawing upon Barad’s (1996, 1997) philosophy of agential realism to rethink the ethico-onto-epistem-ological challenge of involving GenAI in social sciences research, we reveal how human-AI research engagement could inherently be a much more ethical, interdependent, indeed intimate form of collaboration, in which the agency between humans and GenAI needs to be recognised and respected.

In what follows, we first position our study in the broader literature about GenAI in qualitative analysis. We then elaborate on relevant theorising, and how this can add to existing studies about involving GenAI in qualitative research. This includes exploring how post-qualitative ideas could shed light on human-AI collaboration in policy analysis. This is followed by a reflection on the exploratory collaboration in which Danson engaged with Copilot in doing theory-informed policy analysis. This included clarifying the motivation of the collaboration, the relevant policy data, the selection of the GenAI bot Copilot as a research partner, and Bacchi’s (2009, 2012) ‘What’s the problem represented to be (WPR)’ approach as the analytical method informing the analysis of this policy document. Subsequently, we discuss how the human-AI collaborative analytical conversations, which occurred during this process, led to a different way of engaging with GenAI in theory-informed policy analysis, and how this approach triggered Danson to think in ways he had not yet considered – to think the unthinkable. In the final section, we sum up the methodological and ethico-onto-epistem-ological implications of this exploratory human-AI collaboration.

GenAI in qualitative analysis

In recent studies that have deployed GenAI to conduct qualitative analysis, GenAI is predominantly construed as an ‘effective’ research tool for human researchers (Alshami et al., 2023; Jalali and Akhavan, 2024; Lee et al., 2024; Lixandru, 2024; Morgan, 2026; Nguyen-Trung, 2025; Perkins and Roe, 2024; Prescott et al., 2024; Than et al., 2025; Turobov et al., 2024; Wachinger et al., 2024). In this case, the ‘efficiency’ of GenAI is highlighted, based on GenAI’s capability to reduce some of the research workload. Noticeably, most studies (see Alshami et al., 2023; Lee et al., 2024; Morgan, 2026; Nguyen-Trung, 2025; Perkins and Roe, 2024; Prescott et al., 2024; Than et al., 2025; Turobov et al., 2024; Wachinger et al., 2024) followed the methodological steps of thematic analysis to prompt GenAI bots to generate codes and themes to support their preliminary forms of qualitative analysis, with various caveats noted. However, Nguyen and Welch (2026) warn us of the epistemic risks of relying on GenAI to code qualitative data, and argue for a more critical eye in AI-assisted qualitative analysis. In this more critical vein, and working with GenAI in doing theory-informed qualitative analysis, Paulus and Marone (2025) deployed ChatGPT to conduct a discourse analysis to critique the language choices used to describe GenAI in qualitative research on three qualitative data analysis software websites. Building on these studies, we acknowledge the necessity to adopt a more critical approach to engaging with and interpreting theory when collaborating with GenAI in theory-informed data analysis.

(Re)Positioning GenAI in academic research: AI and ethics in action

Engaging with GenAI also foregrounds broader concerns around the ethics of doing so in academic work. Relevant research reported several major ethical issues regarding the output of GenAI, including: plagiarism – synthesising information fragments from various sources without adequate attribution (e.g. Eke, 2023); integrity – claiming the AI output as personal intellectual products (e.g. Perkins, 2023); bias – the reproduction of Eurocentric predispositions and orientations (e.g. Stewart and Zheng, 2024; Zheng and Stewart, 2024) and; fabrication and misinformation – falsifying sources of information (e.g. Balmer, 2023; Currie, 2023).

In many cases, the involvement of GenAI in academic work is broadly characterised as unethical, without substantive regard for either the affordances or constraints of such engagement. In a position paper rejecting ‘using’ GenAI in reflexive qualitative analysis, Jowsey et al. (2025) assert that qualitative research should remain as a distinct human practice to address the ethical and environmental concerns of involving GenAI in qualitative research. However, as pointed out by De Paoli (2026), whether qualitative research should be a distinct human practice is not simply a methodological problem, but also, a philosophical issue. Specifically, our ways of doing (e.g. whether or how GenAI should be involved in qualitative research) are connected with the ways of knowing and being (e.g. what GenAI is and how GenAI should be positioned in social sciences research). To build on De Paoli’s (2026) important critique, our suggestion of a more ‘ethical’, intra-active way of engaging with GenAI by acknowledging both humans’ and GenAI’s capabilities would help facilitate the discussion of current social, environmental affordances, and limitations regarding involving GenAI in qualitative research.

To position GenAI in research, it is necessary to consider the nature of the knowledge (re)produced by GenAI, and accordingly, its role in (re)producing knowledge. Therefore, a close examination of the ethical, ontological, and epistemological considerations of GenAI in social sciences research is essential, and critical post-qualitative perspectives become useful in this regard.

Theorising human-AI collaboration in doing research

Qualitative methodologies are prevalent within social science research. Kelly (2023) argues that qualitative research is often embedded with humanistic assumptions about particular ways of defining sociocultural phenomena and subscribes to particular sets of norms for validity and expectations of outcomes. With the advent of GenAI, ‘traditional’ ways of conducting qualitative research are being challenged, as GenAI enables rapid processing of language patterns that are potentially beneficial for data analysis. This capability of GenAI offers an alternative way to think about doing social science research in a post-qualitative era.

As suggested by Tesar (2020), post-qualitative perspectives challenge the illusion of the exclusive human ‘I’ in producing ‘scientific truth’ through their research. Post-qualitative methodologies offer the possibility to take us on an adventure to seek new ways of researching whilst maintaining the rigour of our research. Doing social research post-qualitatively foregrounds a more-than-human philosophy to guide the design of innovative research methods and challenges the traditional human-centred stance in meaning-making for knowledge production (Tesar, 2020).

In collective writing (Peters et al., 2023), Tesar further discusses the potential impacts of GenAI, predominantly ChatGPT, on social science research by comparing GenAI with challenges brought about by posthumanism. Specifically, Tesar argues that both posthumanism and GenAI challenge human exceptionalism, particularly regarding the binary distinction between human and non-human (Peters et al., 2023). Accordingly, posthumanist studies acknowledge the collaborative nature of knowledge (re)production via the entanglement of human and non-human entities. From a posthumanist perspective, collaborating with GenAI in social science research is to acknowledge the agency between human and AI that could contribute to knowledge (re)production without simply regarding AI as a ‘lifeless’ research tool at human disposal. This posthumanist acknowledgement becomes a crucial consideration for a post-qualitative inquiry involving GenAI in the social sciences.

Barad’s philosophy as an anchor for a post-qualitative inquiry

Given the plural and non-humanistic features of post-qualitative methodologies, Karen Barad’s philosophy provides a useful anchor point for understanding the new materialist and posthumanist standpoint. Barad’s philosophical perspective flags what they describe as the ethico-onto-epistem-ology of research, foregrounding a critical awareness of the entanglement of ontology, epistemology, and ethics (or axiology) in doing research. While ontology and epistemology are usually understood as ways of being and knowing (Barad, 1996, 1997), ethics or axiology is often associated with how knowledge is produced and used (Kelly, 2006).

Barad’s (1997) theorisation of ethico-onto-epistem-ology was predominantly inspired by Michel Foucault’s critique of representationalism, which separates what is represented from the practices of representing. Essentially, a Foucauldian conception of power is central to understanding this critique. For Foucault, knowledge entails power, and power is both repressive and productive (Foucault, 1980a), and only exists in actions (Foucault, 1980b). There is always an ethics involved in such knowledge-power actions – an ethics-in-action. Barad (1997) points out that Foucauldian power is not an external force but ‘an immanent set of force relations that constitutes (but does not fully determine) the subject’ (p.63). From this perspective, the materiality of the body is informed by discursive practices which are infused with a particular ethics-in-action (Barad, 1997). As part of critiquing representationalism, Foucault also became interested in how discursive practices constitute materiality (Barad, 1997). Building on Foucault, Barad (1996, 1997) appeals to the entanglement of ontology, epistemology, and axiology as ethico-onto-epistem-ology to attend to the inherently ethical-materiality produced through qualitative research.

Barad’s (1996, 1997) philosophy of agential realism reflects the nature of such considerations of the ethico-onto-epistem-ological entanglement. Barad’s Foucauldian theorisation of agency is central to understanding agential realism. Barad (1997) argues that Foucauldian power is not deterministic, as the formation of subjects and objects always involves conflict and resistance. From this perspective, the boundaries of defining subjects and objects are always unstable and oriented towards change (see also Haraway’s (1988) earlier argument). Following on from this line of thinking, Barad (1997) extends Foucault’s theorisation of power as a form of intra-activity, rather than the interaction between two clearly-distinguished entities. Accordingly, Barad (1996) argues that agency is not an inherent attribute or feature, but an enactment, or ‘a matter of intra-acting’ (p.183). Agency is the process through which the constructed knowledge becomes embodied as part of our ways of knowing and being, that is, with material consequences (Barad, 1996). Therefore, agency is not ‘possessed’ by human subjects but only exists in the intra-action between temporarily formed subjects and objects. Barad’s (1996, 1997) agential realism thus moves beyond the subject-object distinction to focus on the impact of the intra-action between human and non-human as part of the collaborative meaning-making process.

We also acknowledge and seek to build on important critiques/developments of Barad’s theorisation of agential realism to consider further implications for positioning GenAI in social sciences research. Braunmühl (2017) critiques the overemphasis on the active dimension of the subject/matter and the devaluation of the passive dimension of subjectivity in Barad’s theorisation, as they do not help overcome the subject-object distinction. Similarly, Hollin et al. (2017) extend Barad’s philosophy and argue that agential realism is not just about what is active, but, equally importantly, also about what is absent. These are helpful developments of Barad’s philosophy for thinking about subjectivity in human-AI collaboration. The passive dimension of subjectivity is often connected with conforming, in our case, thinking within the policy logic. It is an important dimension for the realisation of active subject formation relating to resisting, when the unthinkable becomes thinkable, in our case, thinking beyond the policy logic. D’Amato’s (2024) theorisation of GenAI (specifically ChatGPT) as a Foucauldian subject is also helpful here to think beyond the subject-object distinction. While Foucault did not theorise the non-human, D’Amato (2024) argues that ChatGPT demonstrates the features of a speaking, acting subject that is capable of conforming, resisting, and self-forming. From this perspective, our theorisation of intimate, intra-active collaboration with Barad’s philosophy is a process in which both human and GenAI could conform with or resist ideas. Moving beyond human-object collaboration, this form of intra-active collaboration enables humans and GenAI to co-contribute to meaning-making as relatively equal subjects.

Nevertheless, we caution against the sheer positioning of GenAI as an ‘absolute equal’ subject to human researchers, which could be problematic. Calvert-Minor (2013) points out that the dismissal of human centrality in Barad’s philosophy could bring about epistemological challenges, that is, how knowledge is produced and what it is produced for. Building on this development, we argue that the need to acknowledge the capabilities of GenAI does not assume a ‘flattened’ human-AI collaboration. The human researcher should lead the purpose and means of collaborative analysis, which is central to research ethics, while the NLP ‘capacity’ of GenAI to identify, disrupt, and reconstruct language patterns in qualitative data should be recognised and ‘encouraged’.

Thinking with Barad’s philosophy, our ways of being are informed by the knowledge constructed through our ways of doing research. In our case, the ethics of knowing and being are informed by our ways of doing. Specifically, our engagement with GenAI was informed by rethinking their positioning in academic research: Are we using GenAI as a tool to find an answer to our question? Or are we engaging with GenAI to explore alternative ways of knowing and being together – to think differently? What are the material consequences produced through our different ways of doing research with GenAI? What are the ethics imbued within this process?

With the advent of GenAI, the process of prompt engineering between human users and GenAI is a good example of intra-action between these two types of human and non-human capabilities. The focus of the intra-action is on how GenAI might inform, add to, or challenge human thinking of being and knowing. By synthesising Tesar’s (2020) suggestion of philosophy as a method, the anchor point of the post-qualitative human-AI collaboration is to attend to how the meaning-making process during the intra-action between human and GenAI co-contributes to thinking what has not been possible to date in theory-informed policy analysis.

This focus on the intra-action between human and non-human does not completely dismiss the practicality of more traditional qualitative research methods or frameworks. As argued by St. Pierre (2014), post-qualitative inquiry rejects not the conventional qualitative methodology but its underpinning human-centred assumptions. From a feminist perspective on research accountability, Haraway (1988) suggests that the production of knowledge could be tuned to the resonance, that is, what echoes with the researcher’s situated ways of knowing and being. Gerrard et al. (2016) suggest that post-qualitative inquiry would challenge the assumptions of qualitative theories and analytical tools, and also interrogate the ‘scientific’ and ‘objective’ nature of humanistic interpretations of data. From this perspective, the theories and analytical tools that are often used in qualitative analysis could also be adapted to extend the insights provided by more traditional qualitative insights.

(Re)Positioning GenAI in research: Extending collaborative intelligence

While we acknowledge the intricate entanglement of the ethical-onto-epistemology (Barad, 1996, 1997), we also note that the positioning characterising human-AI collaboration alludes to ethical issues in ways that were not previously evident. Crudely, a question often asked in research is: how ethical is it for academics to collaborate with GenAI in their research?

This ethics is deeply intertwined in the onto-epistemology of research and the pragmatics of doing research. John McCarthy (1981), one of the founders of the discipline of AI, distinguishes the epistemological problem of AI from the heuristic problem. The heuristic problem of AI, according to McCarthy (1981), is about the algorithms or the mechanisms of AI to research for information and match or identify particular patterns. The epistemological problem, on the other hand, questions:

what kinds of facts about the world are available to an observer with given opportunities to observe, how these facts can be represented in the memory of a computer, and what rules permit legitimate conclusions to be drawn from these facts (McCarthy, 1981: 459).

GenAI bots, from McCarthy’s perspective, are ‘observational and computational facilities’ (p. 460) that contribute to exploring what is potentially knowable (McCarthy, 1981). Potentially, with the generated responses, GenAI bots do not simply circulate what is already known but could also add to or challenge the thinkable in humans, to inform our ways of knowing and, indeed, ways of being. From a post-qualitative perspective, what McCarthy suggests as the epistemological problem of AI could be further understood as a precursor to the ethico-onto-epistem-ological problem flagged by Barad. The post-qualitative considerations of collaborating with GenAI in doing research should then be centred on the scrutiny of sources, ways of representation of the data, and how the synthesis of language patterns by GenAI may extend (or limit) our ways of thinking.

Resonating with McCarthy’s conceptualisation of the two-fold problems of AI, Russo et al. (2023) suggest a double shift of focus for working with GenAI. As argued by Russo et al. (2023), what humans can trust is not the output of a GenAI bot but their process to generate appropriate outcomes; and the assessment of the GenAI output should always be based on inclusive axiological strategies rather than AI expertise alone. All this occurs within an understanding of the ethics that surround any knowledge development more broadly.

Following on from such ethical concerns, we do not simply deploy a GenAI bot as a research assistant that needs to be trained to maintain the theoretical focus of the analytical approach before involving them in the policy analysis. Instead, we regard GenAI as an integral part of the very agential engagement between the human ‘I’ and non-human ‘GenAI’. In this process, ethics are intimately embedded in human-AI intra-action. From a post-qualitative perspective, this positioning moves away from the humanistic ‘I’ in the meaning-making process in data analysis, instead appealing to recognising and respecting GenAI’s capabilities to challenge or add to human thinking.

This repositioning of GenAI’s role in research also challenges more binary aspects of what is referred to as collaborative intelligence. In much of the literature and theorising, collaborative intelligence refers to the synergy of human intelligence and AI to augment outcomes (see Joseph and Pandey, 2025; Schleiger et al., 2024; Wilson and Daugherty, 2018). In such studies, human-AI collaboration is characterised by a sense of the complementarity between two different types of intelligence (Joseph and Pandey, 2025). As argued by De Cremer and Kasparov (2021), one fundamental difference between human intelligence and AI is that AI is predominantly based on the available data, whereas human intelligence is potentially much more expansive. In what is often called machine logic, AI excels in routine work, such as identifying recurring patterns in the data (De Cremer and Kasparov, 2021; Przegalinska and Triantoro, 2024). Human intelligence, on the other hand, can involve imagination, anticipation, perception, and evaluation of circumstances and situations, with inevitable emotions and power relations impacting reasoning (De Cremer and Kasparov, 2021; Przegalinska and Triantoro, 2024). Human intelligence may also be confined by (powerful) discursive boundaries or conceptual logics produced through particular forms of semiotic signs (Krämer, 2022). These boundaries may delineate what is (temporarily) thinkable to humans.

Much of this literature about collaborative intelligence seems to reinforce the ‘boundary’ between humans and AI. However, we challenge such boundary-making from post-qualitative perspectives. We seek to appropriate a conception of collaborative intelligence which positions AI as a collaborative, agentic partner. Such intra-actions are central to Barad’s understanding of the relations forged through agential realism.

With this repositioning of GenAI, we aim to achieve an intra-active form of human-GenAI collaboration in qualitative research. By avoiding positioning GenAI as an external research ‘tool’ or assistant, we do not ‘use’ GenAI to generate an analytical report. Instead, this intra-active form of collaborative intelligence aims to explore an alternative reading of the data (a policy document) guided by a Foucauldian analytic framework – and involving analysis from both GenAI and human actors. This entails more intimate intra-actions between the human researcher and GenAI to analyse how a particular ‘problem’ is articulated in a specific Chinese education policy.

An example: Danson-Copilot’s collaborative analysis of a Chinese educational policy

In this section, we provide an example of the rich agential intra-action between a human researcher (the first author, Danson) and a GenAI bot (Copilot, powered by GPT 4.0 at the time of collaboration). However, this is not a dualistic analysis that characterises what might be more traditionally (or qualitatively) understood. Instead, Danson’s engagement with Copilot, as the Danson-Copilot collaboration, was a deeply synthesised process of engaging relations between human and GenAI, which revealed both the affordances and limitations of such technologies. Put simply, we seek to elaborate a more post-qualitative positioning and reflection in terms of human-AI collaborative practices.

Setting the scene: The policy

This example is about an exploratory collaboration between Danson and Copilot in doing theory-informed policy analysis. Specifically, Danson collaborated with Copilot to engage with Bacchi’s (2009, 2012) Foucauldian WPR approach to analyse an online Chinese public policy. The focus of this collaboration is not on the heuristic problem – how AI algorithms or coding could be revised or improved to generate ‘better’ analytical responses – but the ethico-onto-epistem-ological problem of the intra-active intimacy of human-GenAI agency in the collaborative meaning-making process.

The online policy chosen as the sample data relates to the initiation of China’s nationwide competency-based education reforms – known as the Suyang (素养) policy. The term Suyang is conceptualised as sinicised competencies (Lin, 2017) and predominantly builds on the Organisation for Economic Co-operation and Development’s (2005) competency framework. The overarching educational goal of this reform is articulated as ‘Lide Shuren (立德树人)’, that is, to cultivate Chinese citizens with good moral character (Ministry of Education, 2014), through the development of core competencies. Within this policy, ¹ there is explicit mention of the need to address the problem of ‘Zhongzhi (重智)’, that is, overemphasis on intellectual development, and ‘Qingde (轻德)’, that is, disregard of moral/character cultivation. ² When written in Chinese, these concepts were placed together to justify the necessity of the development of broader competencies in the Chinese young generations. In a Chinese sentence, there is no space between words, only punctuation. This makes it difficult to ‘decode’ policy assumptions in analysis.

In our policy analysis, the key research question asked was: What educational problems are constructed in the Suyang policy? To investigate this question, we drew upon Bacchi’s (2009, 2012) WPR approach, and applied it to analyse this policy.

Terms of engagement: Why a collaborative human-GenAI policy analysis approach?

To help make sense of this policy as part of an AI-collaboration, instead of conducting a general thematic analysis, we sought to ensure we remained true to the form of intra-active agential realism that characterises a more critical approach to such engagement when doing human-AI policy analysis. In the first instance, we needed to check the extent to which the AI ‘understood’ the theorising we were deploying – in this case, Carol Bacchi’s WPR approach. This approach is compatible with our post-qualitative anchoring point as it also builds on the Foucauldian philosophical perspectives that undergird this exploratory collaboration. This approach also has the emancipatory potential to facilitate human researchers to deconstruct the policy problem by thinking about the assumptions, the unproblematic, and the silenced in policy data (Bacchi, 2009, 2012). Put simply, this approach is also oriented towards thinking the unthinkable, which fits the motivation of our collaboration.

Bacchi (2009) argues that we are not governed through policies but by problematisation. The so-called problems depicted in policy documents are not actually social problems but are constructed as such through policy. The result is a policy presenting ‘contested claims about the existence and nature of social problems’ (Bacchi, 1999: 59). The focus on governing technologies, tactics, and rationales helps dissect how issues are constructed as ‘problematic’ in policy.

Bacchi’s (2009, 2012) WPR analytical approach aims to interrogate the representations, assumptions, the unproblematic, as well as the sites, agents, and effects of the problematisation in policy data. Specifically, Bacchi suggests six sets of questions to guide policy analysis:

1. What’s the ‘problem’ . . . represented to be in a specific policy or policy proposal?

2. What presuppositions or assumptions underpin this representation of the ‘problem’?

3. How has this representation of the ‘problem’ come about?

4. What is left unproblematic in this problem representation? Where are the silences? Can the ‘problem’ be thought about differently?

5. What effects are produced by this representation of the ‘problem’?

6. How/where has this representation of the ‘problem’ been produced, disseminated and defended? How has it been (or could it be) questioned, disrupted, and replaced? (Bacchi, 2012: 21)

Essentially, the WPR approach is underpinned by Foucauldian philosophy which foregrounds how policies, as a state governing technology, regulate our ways of knowing and being.

Before collaborating with Copilot, Danson had conducted a preliminary human version analysis of the chosen policy, drawing upon these insights informed by the WPR approach, but found it difficult to resist how the policy was framing him as its ‘object’, that is, resist being a passive ‘follower’ of the policy as constituted. The fourth analytical question from the WPR approach, in particular, was proving challenging. Consequently, to think beyond what was thinkable to him at the time, Danson thus became motivated to work with a non-human partner to ‘more deeply’ analyse the policy. Upon choosing the GenAI partner, Copilot, Danson conducted an exploratory post-qualitative collaborative policy analysis.

Navigation and analysis in Danson-Copilot’s collaboration

In the collaborative policy analysis, a ‘Navigate and Analyse’ method was adapted from Zheng and Stewart’s (2024) method for prompt engineering to elicit consistent interpretations with GenAI. Based on four moral dilemma story generation experiments with ChatGPT, Zheng and Stewart (2024) highlight the importance of confirming the consistent interpretations of a concept that is central to the desired outcomes between GenAI and human users. In their original two-step ‘Navigation and Generation’ method, they suggested first navigating the GenAI bot to ensure clarification of the concepts or the intended outcomes, and then prompting GenAI to generate text based on the clarifications obtained in the first step. This method is compatible with our human-AI collaboration described here, as it ensures a level of ‘consistency’ in relation to key concepts employed, as well as recognising the capabilities of GenAI in interpreting a concept in a way that may differ from human understanding.

Navigate

In the first step, the navigation was done with Copilot to address two aspects of potential concern. Firstly, Danson started by checking Copilot’s access to the online policy by requesting a summary:

Danson:

Hello! I would like you to analyse a publicly available policy. Here is the link: http://www.moe.gov.cn/srcsite/A26/jcj_kcjcgh/201404/t20140408_167226.html. Could you do a brief summary of this policy for me?

Copilot responded with an English translation of the title of this Chinese policy and a summary of a few key themes in the policy, to prove successful access to the online document.

Secondly, Danson then moved on to navigate Copilot to ensure a clear and consistent positioning of the analytical focus of Bacchi’s WPR approach between them:

Danson:

Thanks for your summary. Next, I will request you to use an analytical approach to analyse this policy. But first I would like you to clarify the analytical approach we are going to use. Could you explain the analytical questions and analytical purposes of Carol Bacchi’s ‘What’s the problem represented to be?’ approach to me?

Copilot responded with some general summaries of the approach. Danson thus added an explicit prompt to obtain more detailed clarifications for a more precise analytical positioning:

Danson:

Thanks! Does Bacchi propose specific analytical questions for this approach? Please also include the references.

This time, Copilot responded with the six analytical questions proposed by Bacchi (2009, 2012), as reviewed above. With the successful navigation of the analytical focus of the WPR approach now underway, Danson engaged Copilot further for a theoretical and philosophical navigation of this approach:

Danson:

Could you then clarify the theoretical and philosophical underpinnings of this approach: (1) Which philosopher’s work did Carol Bacchi draw on to develop this approach? (2) What key concepts underpin the analytical focus of this approach?

Again, Copilot provided some general summaries in relation to poststructuralism and social constructionism without directly answering the first question. Danson then explicitly pointed out the Foucauldian underpinnings of the WPR approach and requested Copilot to provide further clarification:

Danson:

Thanks for the response. I would like to suggest that Bacchi refers to the French philosopher Michel Foucault’s work, such as his theorisation of ‘power’, ‘governmentality’, and ‘problematisation’. Could you explain how Bacchi’s WPR approach is built on Foucault’s concepts?

With this explicit guidance, Copilot responded with consistent interpretations of the theoretical underpinnings as the way Danson understood the approach. In this case, Danson completed the first step of ‘Navigate’ in the collaborative policy analysis.

Analyse

Danson moved on to the second step to prompt Copilot to mobilise the analytical and philosophical clarifications of WPR to analyse the Suyang policy. The analysis was done with several prompts. Danson began by requesting Copilot to apply the six analytical questions to the analysis of the policy:

Danson:

Thanks for your clarifications. Based on the above clarifications, could you use the six analytical questions in Bacchi’s WPR approach to analyse that policy (http://www.moe.gov.cn/srcsite/A26/jcj_kcjcgh/201404/t20140408_167226.html) again? Note: Please include the original Chinese text with an English translation if you want to quote some keywords or extracts from the online policy. Thanks!

Copilot provided some general analytical responses to each of the six questions. Particularly in relation to analytic Question Four in the WPR approach (i.e. What is left unproblematic in this problem representation? Where are the silences? Can the ‘problem’ be thought about differently?), Copilot responded:

Copilot: . . .

• ○ The policy does not explicitly address challenges such as an overemphasis on academic performance or the need for better alignment between different educational levels.

• ○ Silences exist regarding specific implementation strategies and potential obstacles.

To recall, this collaboration sought to better understand how the intra-action between human and GenAI could contribute to new meanings or an alternative reading of the data. Danson thus continued with discussing the analytical responses to the WPR analytic questions, particularly focussing on analytic Question Four which motivated him to take on this collaboration. This is not an easy question, as it challenged Danson to think about what is silenced or unproblematic in the policy. By the time of the collaboration, Danson was uncertain about the analysis he had derived for this particular analytical question in his preliminary ‘human version’ of analysis with the WPR approach.

To further this analytical discussion, and his insights into this approach more broadly, Danson continued by challenging a key point in Copilot’s analytical responses to the fourth question:

Danson:

Thanks! I have a different opinion about your analysis of the fourth analytical question regarding what is left unproblematic. You said ‘The policy does not explicitly address challenges such as an overemphasis on academic performance . . .’. My understanding is that this policy does not justify why the overemphasis on academic performances (重智) would lead to the neglect of morality/virtue development (轻德). Remember the policy proposes to cultivate virtuous citizens, as you specified for the first analytical question. Do you agree or have anything to add to this?

Copilot showed agreement with Danson’s understanding but added a further idea that particularly triggered him to think about what was previously unthinkable to him in his preliminary human analysis:

Copilot:

You raise an important point, and I appreciate your perspective. You’re absolutely right that the policy does not explicitly justify why an overemphasis on academic performance might lead to the neglect of morality and virtue development. The focus on cultivating virtuous citizens implies a balance between intellectual development (智) and moral character (德). However, the policy does not delve into the underlying reasons for potential imbalances or trade-offs.

In this section, the methodological steps taken by Danson in collaboration with Copilot are elaborated. We now turn to reflect on how Copilot’s responses prompted Danson to think differently – to think the unthinkable – during this collaboration, particularly during the final interactions/discussions about the analytical responses.

Thinking the unthinkable: GenAI for collaborative, intra-active policy analysis

In Danson’s preliminary analysis of the Suyang policy with the WPR approach, he wondered, ‘Why should moral cultivation be equally important as intellectual development?’ But this analysis was too simplistic as it did not explain the nuanced strategies underpinning the problematisation focus advocated by Bacchi. Through his collaborative analysis with GenAI, Danson realised how he was being unconsciously governed within the Chinese policy logic/techniques of the Suyang policy. Born, raised, and educated in mainland China, Danson was ‘too comfortable’ with the ways Chinese characters are commonly put together to articulate new concepts in educational policies. For example, in the Suyang policy, the two sets of Chinese characters ‘Zhongzhi (重智)’, that is, overemphasising intellectual development, and ‘Qingde (轻德)’, that is, disregarding moral/character cultivation, were placed together as a new Chinese term to justify the necessity of the development of broader competencies in the Chinese younger generations. Copilot’s responses, in the form of English interpretation, disassembled/disrupted the language patterns in the policy, thereby offering an unexpected perspective to trigger Danson to think further/differently.

Specifically, two aspects became thinkable for Danson, which had not previously been the case. The first ‘lightbulb moment’ happened when Danson began to notice how key Chinese characters were assembled into new, powerful terms to convey the meanings of the policy. As a usual literacy practice, and as a more passive ‘object’ of policy, Danson naturally tried to understand the meaning of the Chinese text by making sense of how different single characters are put together in particular orders when reading the policy. In this process, it is not habitual for Danson, as a Chinese native, to question the ordering or combination of the Chinese characters or why particular characters are selected and assembled into new terms. During the process of constructing the prompt to challenge Copilot’s analytical response to Question Four, Danson became aware that putting together two sets of linguistic characters as a new word, such as putting together ‘Zhongzhi (重智)’ and ‘Qingde (轻德)’, is a powerful process, which does not provide a natural justification for why or how they should be associated. In a Foucauldian sense, the linkage made between these two sets of linguistic characters as two semiotic entities establishes new discursive relations and boundaries (Fairclough, 1992). In the preliminary analysis, before engaging with Copilot, Danson was constituted within a more ‘traditional’, ‘Chinese’ policy logic rather than engaging sufficiently with Bacchi’s problematisation logic, that is, simply re-interpreting the dominant logic of the policy as it stood.

This first lightbulb moment also became possible due to the translation between two languages by Copilot. In this collaboration, Danson communicated with Copilot for collaborative policy analysis in English. As the policy is written in Mandarin Chinese, Copilot’s effective processing of the language patterns inevitably involved a translation process. This translation process by Copilot brought unexpected benefits through disassembling/disrupting the Chinese language patterns that Danson initially ‘took for granted’. Copilot’s translating the policy text from one language to another could be understood as the process of recontextualisation (Bernstein, 1975) where discursive relations represented in one semiotic system were disconnected, revised, and rebuilt into the other.

Second, Copilot’s response prompted Danson to attend to the delineation of knowledge categories when two sets of linguistic characters were compared and contrasted. Copilot’s response in relation to imbalances prompted Danson to question the establishment of boundaries between knowledge categories. In a ‘traditional’ policy logic, emphasising intellectual development and neglecting moral/character cultivation, implies the existence of boundaries between the two categories – intellectual development and moral/character cultivation – of educational development. However, Barad (1996, 1997) argues that the boundaries associated with defining such ‘objects’ are always dynamic and unstable. From this perspective, the possibility of clearly classifying these knowledge categories, which seemed (initially) to be so powerfully established within the policy, is questionable. Danson then noticed the silencing of other possible categories when following this more critical policy logic. This realisation was triggered by Copilot’s response in relation to trade-offs. Even though Danson did not agree with the establishment of boundaries between these categories by the policy, the establishment of such boundaries signifies possibilities beyond intellectual development and moral cultivation in the problematisation. These other possible categories were dismissed in the policy.

We consider this collaborative policy analysis as an inspirational exploratory engagement. In this practice, GenAI Copilot prompted the human researcher, Danson, to interrogate the connections made with the use of language in the policy in ways that were not apparent prior to engaging with AI. This collaboration enabled Danson to develop new understandings beyond the more traditional policy logic with which he was previously struggling. From a Foucauldian perspective, the use of particular language traditions in the example produced more dominant discourses through the policy text, and regulated what was thinkable in human consciousness (Fairclough, 1992). GenAI, empowered by the underpinning LLM, with the ability to translate, dissect, and reconstruct language patterns, sparked inspirations that would not have emerged in a more traditional qualitative approach to policy analysis. GenAI’s capability in deconstructing and reconstructing language patterns, then, disrupted the discursive relations and opened up the possibility for the human researcher, Danson in this study, to think differently.

Essentially, this case is not simply about how GenAI, as an intercoder, generated an answer/code that was unthinkable to human researchers. In qualitative data analysis, intercoder reliability is a common strategy to deploy, involving several researchers/coders cross-checking the analytical results of the same set of data (see MacPhail et al., 2015; O’Connor and Joffe, 2020). Most extant research that incorporated GenAI as a tool or an assistant in thematic analysis (as reviewed earlier) compared human analytical results and AI-generated codes/themes within the data. Their mode of human-AI collaboration predominantly addressed the intercoder reliability from a traditional qualitative research perspective.

Instead, our research reveals a much more integrated, indeed intimate research collaboration, demonstrating the synergy between human intelligence and AI in data analysis. From Barad’s perspective, the unthinkable (to Danson, and also to the second author, Ian, who was intrigued by this exploratory engagement) when GenAI’s capabilities in language dissection and reconstruction intra-acted intimately with human (Danson’s) capabilities. A richly collaborative intra-active series of analytical conversations challenged Danson to reflect on how his thinking was strongly regulated by policy discourses within the Chinese policy context and to re-examine his ways of knowing and being in relation to language use and research engagement. This involved a more spontaneous consensus and interdependence between Danson and AI to achieve a collective synergy.

From the perspective of collaborative intelligence, the intra-action between human intelligence and AI contributed to Danson being able to think what was unthinkable to him initially. This resonates with how human intelligence is not simply ‘inbuilt’ but is an inherently social product, that is, acquired through engagement with others – in this case, GenAI. In people-to-people engagements, human intelligence is inevitably influenced by power relations and the associated emotions produced. In the human-AI collaboration example demonstrated in this study, Copilot’s machine logic effectively identified and disrupted the linguistic patterns of the policy that the Danson initially did not recognise. The outcome was recognition of new discursive relations in the policy text. In the intra-active and more intimate process of collaborative policy analysis, Danson’s human reasoning, informed by the theory, was engaged by Copilot, which enabled him to examine the policy logic regarding what was assumed and silenced. Importantly, his human affective and intuitive awareness enabled him to reflect on: how his initial struggle over policy analysis was connected with the governing techniques of the Suyang policy, and how he may challenge the powerful policy logics which constrained his initial thinking.

Previously, Danson was engaging with the passive dimension of his subjectivity to conform with the policy logic, resembling also a policy ‘object’. Through intra-active engagement with GenAI, he became a much more active agent to resist the policy logic within the policy taking space. As the ‘object’ of policy, Danson was ‘trapped’ by the policy logic as constituted in more traditional rationalistic and ‘magisterial’ interpretations. However, as the forming ‘subject’ within the policy making/taking space, he adopted a more agentic position and was able to think beyond his initial interpretations of the policy as presented. The collaboration with GenAI enabled this to happen.

Conclusion

In this paper, we report a different way of engaging with GenAI in doing theory-informed policy analysis. Building on Barad’s (1996, 1997) philosophy of agential realism, we argue that our human ways of knowing and being are closely intertwined with the ethics of positioning and working with GenAI. Such theorising enables us to explore how more respectful intra-actions recognise the agency between humans and GenAI, and how such recognition might contribute to an alternative way of thinking about data analysis. Guided by Barad’s posthumanist line of thinking, we regard GenAI as an integral part of an agential engagement between the human ‘I’ and the non-human ‘GenAI’, which incorporates a much more considered and robust ethics of engagement, in these relations. In our collaborative experiment, an adapted ‘Navigate and Analyse’ prompting engineering method also proved to be effective in guiding human-AI collaborative policy analysis.

Through the policy example presented, we illustrated how considering human-AI collaboration with post-qualitative ideas could enable the human researcher to think the unthinkable, that is, to think beyond the discursive boundaries established by more dominant policy logics that confined what was thinkable to Danson alone. We position this intra-active human-AI collaboration as a more agentic, critical form of collaborative intelligence. Such repositioning also responds to broader discussions about the ethical concerns of engaging with GenAI in academic research. Concerns about plagiarism, integrity, bias, and fabrication of information reflect more acritical engagement with AI outputs. A more intimate, intra-active, collaborative intelligent approach fosters a much more robust analytical process that moves beyond simply ‘copying’ information or simply ‘relying’ on the output generated by a GenAI bot. Aligned with the onto-epistemological problem of AI (McCarthy, 1981; Russo et al., 2023), we argue that we can benefit from the analytical conversational collaboration process between AI and human intelligence, rather than simply the AI-generated output per se. In this intra-active process, as evident in our example, GenAI could disassemble and disrupt the taken-for-granted linguistic patterns – the realisation of powerful discursive relations – to prompt humans to think beyond discursive or conceptual boundaries.

By not simply using GenAI to generate an answer but engaging with GenAI in analytical conversations, human-nonhuman agency is in an integral and entangled process which makes possible the co-creation of new meanings in data analysis. The exploratory, intra-active, agential engagement is also an instance of a different way of doing, that is, positioning GenAI as an intimate research partner rather than a tool, and reveals how such relations might produce new ways of knowing and being as researchers.