I ended my last entry (“‘Engaging’ with generative AI”, 28 June 2026) by asking if a “reflexive” GenAI is feasible and what would this look like. My query was prompted by Zheng and Hardy (2026) who use GenAI (Copilot) in their WPR study of a Chinese education policy document pertaining to the initiation of national competency-based education reforms.
In that entry I pointed out that Zheng and Hardy omitted a part of the WPR approach, specifically the undertaking to apply the WPR questions to your own problem representations. This undertaking is named as Step 7 of the approach in Bacchi and Goodwin, 2016, p. 20, and as Process 7 in Bacchi 2026, p. 24. It is intended to encourage “self”-problematisation, explained in some detail in Chapter 6 of Bacchi 2026. I note in the 2026 book that I changed the wording from “Step 7” to “Process 7” to reinforce the interconnected nature of the WPR forms of analysis (Bacchi 2026, p. 25 fn 2). ChatGPT tends to refer to Step 7.
At the end of the previous entry on Zheng and Hardy (28 June 2026), I was left wondering what would have happened to the analysis if Process 7 has been included – how would GenAI deal with “self”-problematisation? I decided to find out.
Stage 1: Applying the WPR template
I borrowed the policy document from Zheng and Hardy and prompted ChatGPT: “Apply the seven questions in Bacchi’s WPR to China’s education policy, called the Suyang (素养) policy.” (In this entry I am restricting my comments to experiments with ChatGPT-5.5.) Note that I say “seven questions” to ensure that Process 7 is included.
ChatGPT responded by clarifying its task: “I’ll treat ‘Suyang/素养’ as the Chinese education-policy discourse around student ‘core competencies’ or ‘quality/competence cultivation,’ then map it through Bacchi’s seven WPR questions”. Note the reference to “seven” questions.
To begin, the ChatGPT analysis stated that “Below is one way to apply Bacchi’s ‘What’s the Problem Represented to be?’ (WPR) framework to China’s Suyang / 素养 policy, usually translated as core competencies, key competencies, or competence-based education”. Allow me to highlight the reference here to “one way” to apply WPR, linked to the proposition considered later in this entry that there are different “versions” of WPR (see below).
ChatGPT proceeds to apply WPR, addressing each question and the undertaking to apply the questions to your own problem representations (i.e. Step/Process 7). For those of you who have engaged this program previously, you will not be surprised to find out that it does a superlative job. I could not help but be impressed by the insights generated.
I do not intend to run through the complete analysis as you can access this material yourself by asking ChatGPT to apply WPR to the selected Chinese language policy (you may of course receive modified analysis). Instead, I intend to zoom in on the topic of interest in this Research Hub entry – how did ChatGPT deal with Step/Process 7 and hence with “self”-problematisation?
Stage 2: Applying Step/Process 7
ChatGPT introduces this task by asking analysts “to turn the lens back on themselves”. It proceeds to produce some perceptive insights. For example, it notes:
“In my analysis above, I have represented the suyang policy partly as a form of governance through competence. That framing carries its own assumptions. It may overemphasise control, standardisation, and subject formation, while underplaying genuine educational aspirations within the policy: reducing rote learning, broadening student development, supporting moral education, and making learning more relevant. (Analysis performed using ChatGPT on 18 May 2026; emphasis in original).”
ChatGPT proceeds to note that the “critical policy tradition” “implicitly drawn on” in the analysis “risks treating policy ideals as mainly disciplinary or instrumental”. It therefore concludes that “A fairer analysis would hold both possibilities together”:
“suyang policy may be both an attempt to humanise education beyond exam scores and a way of expanding the state’s capacity to define and monitor desirable student subjectivities. (emphasis added)”
Here I want to highlight the phrase “A fairer analysis”. The language of fair and unfair does not do justice to the “self”-problematisation undertaking. The point is not to pass judgement on your problematisations but to probe them for unexamined assumptions. As Stengers (2008, p. 46) explains, there is a need to “make ‘us’ hesitate about our own conditions of thought” (pp. 41–42). Here Stengers is signalling the need to pay attention to the deep-seated epistemological and ontological assumptions that make “our own conditions of thought” possible. In addition, the statement that a more desirable analysis (“a fairer analysis”) would hold the two propositions together (“hold both possibilities together”) imposes a “neutrality” that has political effects – as I proceed to explain.
ChatGPT ends the section on Step/Process 7 by describing “my own problem representation”:
“The suyang policy may transform the apparent problem of narrow education into a broader project of governing students, teachers, and schools through competency discourse.”
It then offers an “alternative reading”:
“The suyang policy is an attempt to reconcile Chinese moral-political educational traditions, global competency-based reform, and practical dissatisfaction with rote learning.”
Importantly, the ChatGPT analysis concludes that “both readings are plausible”: “A strong WPR analysis would not simply decide which one is ‘true,’ but examine how each framing makes some things visible and others less visible”.
I took exception to the suggestion that both “readings” were “plausible”. Indeed, the reference to the contrasting problem representations as “readings” reveals an approach to WPR distinctly at odds with its development in my writing. In that work the goal is to consider how governing takes place through selected policies and governmental mechanisms, not to produce competing accounts of modes of governance. I decided to pursue the point with ChatGPT.
Prompting ChatGPT
I prompted: “Doesn’t your argument that both readings are plausible ignore the political dimension of the issue?”
With this prompt, things started to get interesting.
ChatGPT responded: “Yes – that is a fair criticism”.
I was happy to hear it.
It continued:
“Saying ‘both readings are plausible’ risks making the analysis sound overly balanced or neutral, when WPR is specifically designed to expose how policy problematisations are political. In this case, the politics are not incidental; they are central.”
I was very happy to hear it.
We are then offered what is described as “a better WPR reading” and “a more critical Bacchian version”, “a sharper version” and a “stronger WPR analysis”. The “sharper version” would conclude: “The suyang policy should not be understood as a neutral correction to exam-oriented schooling. It is a political problematisation of the student, the teacher, and the purpose of schooling.”
As with the language of “fairer” above, the suggestion that there are “stronger” and “weaker” versions of WPR misses the point and purpose of WPR as an analytic strategy. Its goal is to encourage political reflection on governing practices in order to highlight where changes may be needed. A WPR way of thinking provides a framework for reflecting critically on these variables. Such critical reflections “sow the seeds of judgement” (Osborne 1998 in Rose 2000, p. 59), helping to make judgement possible.
By contrast, in this application of WPR, the analysis is put forward as balanced or neutral, in line with ChatGPT’s insistence that it does not engage in politically partisan issues, as pointed out in the earlier Research Hub entry 29 Sept.2023 on the topic. ChatGPT produces analyses in line with the theoretical framing offered by WPR but it does not itself take a political stand. It suggests the situation is a matter of “fair” or “less fair” analysis, or of “weaker” or “stronger” views.
Note, I am not saying that ChatGPT is politically neutral. In the earlier entry (29 Sept 2023) I point to a neoliberal bias in the model I used. Here, I am highlighting that ChatGPT claims to be neutral. I would argue that the position of “neutrality” that it defends is a political stance! Referring to “weaker” and “stronger” versions of WPR neutralizes the political analysis – very much a political effect that requires questioning.
What does all this mean in terms of how we can use ChatGPT? In the previous entry (28 June 2026) I highlight its usefulness in translation. I would also suggest that the analysis performed in this particular case is very useful and would certainly alert readers to points that might not have occurred to them. Still, it remains important, I believe, to recognise the depoliticizing effects of its defence of political neutrality.
The politics of “prompting” WPR
The point I make above is driven home if we consider the kind of analysis produced prior to my second prompt (repeated here from above) -“Doesn’t your argument that both readings are plausible ignore the political dimension of the issue?”.
You may recall that the conclusion ChatGPT reached prior to my second prompt is that two opposed readings of China’s education were plausible. It reads:
“The suyang policy may transform the apparent problem of narrow education into a broader project of governing students, teachers, and schools through competency discourse.”
It then offers an “alternative reading”.
“The suyang policy is an attempt to reconcile Chinese moral-political educational traditions, global competency-based reform, and practical dissatisfaction with rote learning.”
Without my second prompt, the matter would be left there, with researchers more or less being invited to choose one or the other proposition. Underlying political issues are left undiscussed, encouraging a poorly developed understanding.
I raised the same issue in the earlier Research Hub entry (29 Sept 2023). There I addressed the question of whether or not AI has a left-leaning bias, a common characterisation (see earlier entry). By examining the responses to the WPR questions, I was able to show a neoliberal or liberal leaning rather than a structural perspective. None of this would have emerged without useful prompts. As I conclude in that entry: “it was possible to shift ChatGPT in the direction of a more structural analysis through asking explicit follow-up questions”. As I go on to argue, to pose such follow-up questions requires a prior engagement with the literature and with contesting points of view.
In the case at hand, without the second prompt, we would be left with a view that WPR can either support the status quo or not – it’s up to you. It is argued: you can “choose” a weaker or a stronger version of WPR.
This result suggests to me that we need coaching in how to prompt. Basically, what is reported reflects what is asked. We need to find ways to ensure that politics remains front and centre of the questioning. With my second prompt (see above) I managed to extract a more structural understanding of the “problem” of Chinese education policy. Could I have pushed the matter further? Possibly.
The key outcome, as I see it, is to recognize that the material produced in ChatGPT reflects the material available. There is an inherent conservative bias. When or if we ask it to apply WPR we can expect something like what is produced in the reading above – a “balanced” view of the matter. The final message, then, is to tread cautiously around the analysis produced.
“Politics of the prompt” (Amoore 2026)
An important article by Amoore et al. (2026) on the “politics of the prompt” allows us to understand what happened to WPR at the hands of ChatGPT – how it was neutralised. They describe “the prompt” as a “mutation in the exercise of political power”: “the formulation of the prompt changes the politics of what can be said, what questions can be asked, what decisions can be made”. As I outlined in the previous section, my second prompt made it possible to produce a more structural analysis of the Chinese education reforms. Without that prompt, the analysis would have been politically “neutral” and hence neutralized.
Amoore et al. (2026; Abstract) locate themselves within a tradition of thought that conceives of government as “the broad ensemble of ‘techniques and procedures by which one sets about conducting the conduct of self and others’ (Foucault 2010, p. 4).” Working in that governmentality tradition (Bacchi 2026, Chapter 7), they describe “the prompt” as
“an emerging political technology which makes possible a wider governmental rationality – a means of conducting conduct, governing population and ordering the space of what can count as the political (Barry, 2001; Dean, 1999; Miller & Rose, 1990) Amoore et al. 2026, p. 575, emphasis in original).
The authors identify three defining features of the “governmental rationality” of “the prompt”: describing a task, eliciting a desired response, and engineering a prompt. When describing a task, “the user’s interaction with the model is always also the shaping of the model and the world as such” (Amoore et al. 2026, p. 581). Eliciting prompts involves eliciting a desired behaviour from the model (p. 583). To prompt heralds the arrival of the “prompt engineer”, “a novel form of expertise and a new kind of self-governing subject” (p. 586).
According to Amoore et al. (2026, p. 582) “the prompt frames, delimits and condenses what can be known of the problem (‘what do we know about?’) and aligns that problem ever closer to the task that best fits the model’s view of the world.” Here I’d point out how references to “problems” and “problem-solving” form part of the taken-for-granted framework of “computational thinking” (Bacchi 2026, pp. 206-208). In this thinking, the goal is to find the most appropriate prompt “to allow a language model to solve the task at hand” (p. 586). There is no space in this thinking to query the nature or purpose of the task. The status quo is reinforced.
If “prompts” shape social relations, we need to attend to them carefully. It’s all in the questions. You will get what you ask for. As Baum and Villasenor (2023; emphasis added) note,
“seemingly small changes in how a prompt is constructed can lead to very different responses. This is because AI-powered chatbots identify which data to draw from in a manner that is highly sensitive to the specific phrasing of the query”.
To repeat the point made in the earlier entry (29 Sept 2023), thinking precedes prompting. I’m not suggesting here that we become “prompt engineers” (see Amoore et al. on this topic). Rather, we need to become aware of the political dimension lurking within apparently “neutral” commentary, and consider how your prompts can affect that analysis. The initial ChatGPT output needs to be approached sceptically with follow-up questions that ensure politics remains a primary consideration. Beyond recognizing how prompts shape realities, the politics of the prompt includes consideration of the effects that accompany a failure to prompt.
Conclusion
As indicated above, I was deeply impressed with many of the insights generated when I asked ChatGPT to analyse China’s education reforms using WPR. My specific goal was to observe how Step/Process 7 was handled. The section on Step/Process 7 displayed some useful consideration of the assumptions within ChatGPT’s WPR analysis. Things “broke down”, I would say, in the (1) general framing of WPR as “strong” or “weak”, as appearing in “versions”, and (2) the suggestion that blending these “versions” is a useful way to proceed. Here ChatGPT fails to apply “self”-problematisation to this analysis and hence fails to recognize the politically quiescent view of social relations that accompanies it.
References
Amoore, L., Bennett, SJ, Campolo, A., Jacobsen, B. and Rella, L. 2026. Politics of the prompt: Government in the age of generative AI. Economy and Society, 54:3, 573-596, DOI: 10.1080/03085147.2025.2560177
Bacchi, C. 2026. What’s the Problem Represented to be? A new thinking paradigm. NY: Routledge.
Bacchi, C. and Goodwin, S. 2016. Poststructural Policy Analysis: A Guide to Practice. Second edition, 2025. NY: Palgrave Macmillan.
Barry, A. (2001). Political machines: Governing a technological society.Bloomsbury.
Baum, J. and Villasenor, J. (2023). The politics of AI: ChatGPT and political bias. Brookings, 8 May. (https://www.brookings.edu/blog/techtank/2023/05/08/the-politics-of-ai-chatgpt-and-political-bias/
Dean, M. (1999). Governmentality: Power and rule in modern society. Sage.
Foucault, M. (2010). The government of self and others: Lectures at the Collège de France 1982–1983. Palgrave Macmillan.
Miller, P. & Rose, N. (1990). Governing economic life. Economy and Society, 19(1), 1–31.
Rose, N 2000, Powers of freedom: reframing political thought, Cambridge University Press, Cambridge. DOI: 10.1017/CBO9780511488856.
Stengers, I 2008, ‘Experimenting with refrains: subjectivity and the challenge of escaping modern dualism’, Subjectivity, vol. 22, pp. 38–59. DOI: 10.1057/sub. 2008.6.
Zheng, D. and Hardy, I. 2026. “Thinking the unthinkable: Generative AI for collaborative, intra-active policy analysis”. Methodological Innovations,
Carol Bacchi 