I began preparing for my first Oxford seminar on science and religion with a fairly ordinary task. Dr Shaun Henson’s reading list asked us to assess typological approaches, compare their strengths and weaknesses, and consider alternatives of our own. Ian Barbour’s familiar categories—conflict, independence, dialogue and integration—provided the starting point. I expected to read the arguments, decide which framework seemed most convincing, and arrive ready to discuss it.
Instead, I kept interrupting the preparation with a question that seemed more elementary than the assignment: what problem were these classifications supposed to help me understand? I could repeat the four terms before I could explain why arranging the relationship in this way mattered. When I brought my own experience of AI-assisted theological writing into the discussion, the difficulty became sharper. I usually began with a question, a text or something I wanted to make. Computing and theology already participated in the activity. I had rarely felt obliged to reconcile them before starting.
My first response was to suspect that AI and theology belonged to a different paradigm. Perhaps these classifications felt unfamiliar because my work had already moved beyond the situation they described. But there was another possibility: perhaps I had not yet understood the authors well enough to identify the difference. I could not initially separate a limitation in the frameworks from a limitation in the explanations I was receiving—or in my own reading.
The preparation consequently developed through repeated excursions and returns. We moved into my earlier writing, the purposes of learning, asymmetry, engineering, cybernetics and the changing capabilities of AI. Several times I explicitly brought the discussion back to Henson’s readings and the task of assessing them. Each return asked something practical of the excursion: could I now explain more clearly what a typology accomplished, defend a comparison between approaches, or formulate an alternative worth investigating? Sometimes I could; sometimes an attractive connection had taken us away from the question I had actually asked.
Looking back, this movement is the organising feature of the inquiry. The wider questions helped me understand the assigned problem, while the assigned problem gave me a reason to test the wider questions. I did not begin with a settled method for doing this. Its shape became visible through the repeated need to refocus—and through discovering that I returned to the same readings with different expectations.
I needed the problem before the classification
The AI’s first answer was organised and apparently helpful: a comparison of the authors, possible questions for the seminar, and a provisional preference for Stenmark. Yet I had to ask for an entry-level explanation. Being told which framework offered greater analytical detail did not help enough when I still could not see what that detail was for. The preference was also the AI’s assessment. I had not yet reached it myself.
The explanation became more concrete through an example: an evolutionary account of human development placed beside a Christian claim that human beings are created by God. Whether these claims contradict one another depends partly on what creation means in the particular account and what further conclusions are being drawn from evolution. Barbour’s categories then acquired a purpose. Conflict identifies incompatibility; independence separates domains or functions; dialogue explores connections while retaining distinctions; integration seeks a more unified account (Barbour, 2002).
This helped me understand the classification, but my own experience produced another question. In my writing, AI and theology usually seemed to help one another. Did the absence of an initial conflict make the framework unnecessary? The AI pointed out that Barbour already included constructive relationships. That correction was reasonable. It still left me wondering why I should begin by deciding which relationship I had, rather than following what happened while I worked.
When I again asked what lay underneath the readings, the AI suggested that they could feel abrupt because they discussed how to organise a question before making clear why anyone needed to ask it. This formulation redirected my attention to the order of learning. I immediately pressed the point: why ask us to evaluate ways of organising a problem before we understand the problem? The AI then qualified its explanation. A reading list beginning with typologies did not establish that the authors themselves omitted the background. We were entering chapters and articles within an existing debate; the seminar had not yet taken place. My difficulty was real, but it could not tell me what the lecturer assumed.
Only then did the AI’s proposed sequence become useful to me: encounter something puzzling, identify what needs explaining, introduce frameworks, and assess whether they clarify the case. I recognised a way to connect the assignment with my own habits of inquiry. The sequence emerged from our exchange; it was not a method I had found in Barbour or a named educational theory I could claim as my own.
Adam Chin’s article gave that concern an existing scholarly location. He distinguishes classifying possible science–religion relationships from classifying scholars and their work. These activities can serve different purposes, even when they use similar labels. His challenge to the usefulness of categorising authors is unusually direct:
“But what can a scholar (or a lay reader) do with that kind of information?” (Chin, 2023, p. 663)
This question mattered because I had been treating my difficulty as a failure to understand an established framework. Chin showed that its practical intellectual purpose was itself under discussion. His alternative groups scholarship through methods—conceptual analysis, case studies, relativising approaches and fieldwork—so that a reader’s concern can guide the choice of investigation. This is a classification of scholarly work, with a different object from Barbour’s classification of relationships. Calling one “better” without identifying the task would conceal that difference (Chin, 2023, pp. 656–657, 668–674).
I then asked whether this process of finding the question could itself belong in my seminar contribution. Until that point, I had treated my confusion mainly as something to overcome before the real discussion began. It was becoming a source of criteria for the discussion: what does a framework enable us to notice, distinguish or investigate? Orienting a beginner, clarifying an inference and explaining a historical encounter are different achievements. I needed to know which one I was asking a typology to deliver.
When I returned to the seminar questions at this stage, I wanted the path into the material to be part of what I could share. I described it as bridging the gap between receiving the reading list and making it my own inquiry. I also insisted that the AI distinguish the original authors’ arguments, ideas already present in my writing, and connections it was proposing. Otherwise, a persuasive synthesis could make an established argument look like my discovery, or turn an AI suggestion into a position I had supposedly accepted. Keeping these origins visible was becoming part of the method.
My objection had to survive the original authors
I initially resisted the four categories because they seemed too mechanical for actual lives. My first response was almost combinatorial: if several relationships could occur together, why restrict the description to four types? A person might disagree about one claim, cooperate on another task and leave a third question open. But even combining categories seemed insufficient. I began insisting on an iterative process: an exchange changes what a participant thinks, and that change affects the next exchange. I was moving from the number of possible combinations towards the history of their formation.
The return to Barbour required a correction. In the original book, before developing the four categories, he writes: “Particular authors may not fall neatly under any one heading”; he immediately adds that agreement on some issues can coexist with disagreement on others (Barbour, 1997, p. 77). His later response to critics also acknowledges complex historical interactions and the need to revisit evidence. My criticism could therefore not fairly proceed as though he had stipulated four exclusive containers for every human encounter. Counting combinations would not solve the difficulty either: these categories are not independent mathematical switches, and ordinary conversation is not identical to “dialogue” in his technical usage.
This clarification did not immediately satisfy me. Barbour’s defence includes the analogy that “A guidebook to any territory is not intended as a substitute for firsthand exploration” (Barbour, 2002, p. 348). I understood the point, but wondered what would count against the guide’s usefulness. If complicated cases could always be accommodated by invoking selectivity, how would I distinguish a helpful simplification from a misleading one? My frustration became explicit when the discussion continued to speak broadly of science and religion. Which science—physics, biology, computing? Which religion, and which aspect of it? Selective scope could be legitimate; the usefulness of the selection still needed demonstration.
Geoffrey Cantor and Chris Kenny had already made historical particularity central to their criticism. Their discussion of the comparative anatomist St George Jackson Mivart showed why it mattered. They describe him criticising Darwinian explanations of human mental and moral capacities, defending disciplinary independence when discussing Galileo, recognising rational inquiry in both science and religion, and pursuing research informed by a divine architect. The relationships differed with the issue. The historians’ interpretive point was that “the individual must be treated as an active agent who deploys different strategies creatively” (Cantor and Kenny, 2001, pp. 777–778).
This challenged something in my own response. I had contrasted “historical encounters” with the complexity of a person’s real life, as though history necessarily offered the thinner account. Their proposal makes biography a major site of investigation precisely because commitments develop through particular experiences and changing circumstances (Cantor and Kenny, 2001, p. 779). My concern for lived complexity connected with a method already present in the readings. The disagreement with Barbour then became more interesting: could his categories help reveal that complexity, as he maintained, or did organising the account around them obscure what made the person’s activity intelligible?
Stenmark seemed closer to what I wanted because his dimensions allowed a more differentiated description. Social organisation concerns participation, institutions and transmission; teleology concerns purposes; epistemology concerns reasoning and justification; theoretical content concerns what the accounts are about. Two practices could cooperate in one respect and diverge in another. Yet I asked whether these aspects could interact and change, perhaps even generating something that the initial description had not anticipated. If the framework omitted that possibility, might this be an opening for my contribution?
The original text removed the easiest claim to novelty. Stenmark describes science and religion as “dynamic and evolving social practices,” and explicitly denies that their boundary can be fixed in advance. He proposes “at least four levels or dimensions,” leaving the number open (Stenmark, 2004, p. 12). He then explains that “the answer one gives with respect to one of these dimensions often has consequences for one’s conception of the other dimensions as well” (Stenmark, 2004, p. 15). Change and interaction were already there. The remaining question concerned what the framework enabled us to investigate about a particular transformation.
Andrew Loke’s newer fourfold taxonomy supplied another reason to revise my first objection. Its categories—Conflict, Compartmentalization, Conversation and Convergence—classify perceived or expressed relations between specified aspects of science and religion. Convergence concerns a perceived relation of evidential support. The improvement he proposes therefore depends on identifying the relevant claims and perceptions, rather than simply multiplying the boxes (Loke, 2023, pp. 31–35). I could still ask whether this helped explain a developing encounter, but the number four was no longer a sufficient criticism.
This was an important correction in the AI-assisted process itself. I had sometimes been reacting to a compressed representation of an author. The AI helped articulate the objection, but it also helped produce the simplified target. Returning to the original passages changed what I could responsibly criticise. It would be misleading to narrate this exchange as a sequence in which my intuitions were repeatedly confirmed.
Two passages nevertheless preserved the force of my concern. Barbour states that his typology concerns “fundamental science as a form of knowledge,” distinguishing that purpose from the study of applied science and technology, whose ethical significance he addresses elsewhere (Barbour, 2002, p. 352). Stenmark’s concluding discussion is especially revealing. Religious influence on scientific direction, development and application is difficult to represent through relations among goals, methods and theoretical outputs. These activities involve social processes, and he acknowledges:
“I do not know at the present time how to do this in an elegant and illuminating way, and on this issue we certainly need help from the historians and sociologists who are engaged in the science-religion dialogue.” (Stenmark, 2004, p. 267)
This is more specific than saying that Stenmark forgot the social dimension: he introduces it near the beginning and returns to the difficulty of incorporating it adequately at the end. Acknowledging a dimension and providing an illuminating account of its operation are different achievements. That distinction gave my criticism a firmer textual basis.
My first reaction to Barbour’s restriction was another objection: applications are integral to science, especially to the computing I practise, so how could they be excluded? The AI distinguished a limited inquiry from a defective one. I did not have to accept the restriction as adequate for my project in order to recognise that distinction. Designing a system, selecting sources and evaluating its use are central to my encounter with AI. A framework focused on relations between knowledge claims could still be useful, while leaving those activities insufficiently examined. This was a more precise reason to seek additional methods.
I returned to the assignment with a changed standard of comparison. I could no longer contrast my dynamic account with uniformly static predecessors. I needed to ask what each approach selected for attention and what additional work its distinctions made possible. Describing a relationship, explaining how it arose, and judging how it ought to change require different arguments. A framework might contribute to all three without accomplishing any of them merely by naming a category.
Who actually needs reconciliation?
The question of asymmetry developed through another answer that initially seemed relevant but did not reach my difficulty. I asked whether reconciliation was often something religion sought with science, while scientific work had no corresponding need. The AI translated this into a question about who should change. It explained positions giving priority to science, to religion, or to a case-by-case judgment. That distinguished possible responses to disagreement, but I returned to the objection: the participants might be addressing entirely different issues.
Consider a hypothetical scientist investigating an empirical question and a theologian interested in its implications for creation. The scientist may have no professional reason to pursue that theological implication. The theologian may nevertheless regard it as consequential. Their interests differ before any disagreement has been identified. The scientist’s lack of interest would describe a particular professional situation; a philosophical commitment to the independence of science and religion would make a further claim.
I was also concerned that common language could hide different activities. An engineer and a theologian might both speak about learning while referring, respectively, to a training process and the formation of a person’s judgment. They might use the word purpose for a system’s assigned objective and for an account of what a life is for. These were examples of the problem I wanted to identify, not claims that the disciplines possess fixed, incompatible vocabularies. Before comparing their answers, I needed to know whether they were answering the same question.
The AI then reformulated the problem: before asking which side should change, we must establish whether there is a shared disagreement and who regards it as a problem. Returning to Stenmark clarified the limit of the criticism. His discussion begins, “Suppose that there really is a conflict,” before considering whether religious beliefs, scientific theories, or neither should automatically take precedence (Stenmark, 2004, p. 266). I was asking about a prior condition: how something becomes a shared issue at all. A rule for judging a conflict does not establish that the participants perceive one.
The additional original texts make this concern less isolated. Drees opens his book with a lecture encounter in which his scientific account of natural history and an audience member’s theological question about the Fall did not initially meet. His conclusion is concise: “Debates are often non-debates, as issues and criteria are framed differently by the various participants” (Drees, 2010, p. 1). The example matters because it locates the difficulty in an actual exchange. It does not require either participant to be unintelligent or insincere.
John H. Evans and Michael S. Evans offer a related methodological challenge from sociology. They argue that assuming science–religion conflict is necessarily about competing ways of establishing truth can conceal other empirical possibilities. Their review directs attention to institutions, interests and public disputes as well as systems of ideas (Evans and Evans, 2008, pp. 87–91, 98–101). This does not establish my suggestion that most encounters are asymmetric. It gives a reason to investigate what is at stake before deciding what kind of disagreement we have.
I was tempted to generalise this asymmetry across the field. Our discussion did not supply the evidence to establish how often it occurs. It did give me a question worth investigating. The wish to relate two activities belongs to particular people, institutions and projects; it should not be attributed automatically to everyone whose work enters the discussion.
My own experience introduced a different asymmetry. AI could help articulate an argument, suggest an objection or reorganise a draft without sharing my theological commitments. Its responsiveness could make the work feel unusually smooth. That experience was real, but it did not establish the truth of the argument or demonstrate agreement between two entire disciplines. I remained responsible for checking the sources, judging the interpretation and deciding what to publish.
At one point I suggested that alignment might need to come before constructive dialogue. That formulation also required unpacking. Participants need enough coordination to begin—some way to understand the task and recognise a relevant contribution. They do not necessarily need agreement about ultimate commitments, and the conversation may help establish the coordination that was initially missing. In my AI work, a responsive output could support the activity without establishing theological agreement. Cooperation became something to investigate: what enabled it, what each contribution amounted to, and whether its apparent smoothness concealed differences in purpose or judgment.
Returning again to the readings, I could see why Henson’s plural “Sciences and Religions” mattered. A useful account had to identify the practices and participants involved, including differences within a discipline or tradition. My questions about asymmetry and language had brought me back to the assignment with a more demanding understanding of its object.
A historical precedent cannot decide present adequacy
A later return to the original authors began with another practical question: were their examples mainly biological? Several examples the AI had offered concerned evolution, but that selection did not establish the range of the authors’ work. A separate survey-visualisation project also entered the discussion. When the response continued along that line, I explicitly set it aside: I was asking which sciences the original authors themselves examined. This was one occasion when refocusing required leaving a relevant-looking topic alone.
The books corrected the impression. Barbour’s treatment includes quantum theory, relativity and cosmology as well as evolution; Stenmark’s 2004 book gives evolutionary debates a particularly prominent place. Barbour even suggests that Dialogue and Integration appeared more common among physicists and cosmologists than biologists, while identifying books and conferences as the basis for that observation (Barbour, 1997, chapters 7–10); (Barbour, 2002, p. 350). This was an author’s assessment, not a systematic comparison. The question I could defend concerned how frameworks developed through these subjects travelled into computing and engineering.
Checking the original texts for this revision adds a useful complication. Amy Lee’s language-based approach treats science and religion as developing social discourses that shape perception. Yet she explicitly excludes “engineering, medicine, computer science, or the social sciences” from the chapter’s use of science and concentrates on Christianity (Lee, 2022, p. 30). Her conceptual approach may still inform my project, but applying it to AI-assisted theology requires an extension. Even an alternative to typology has a scope that needs examination.
The sharpest disagreement with the AI came when I suggested that contemporary AI might require approaches different from those developed through familiar discussions of evolution, physics and cosmology. It responded with an important historical correction. In 1998, Anne Foerst had already examined the humanoid robot Cog in relation to theology, including the image of God, and proposed a framework for mutual enrichment. Here was an earlier engagement with a concrete AI project, not simply an abstract speculation about future machines (Foerst, 1998).
The example challenged any suggestion that theological engagement with concrete AI projects had begun recently. I nevertheless pushed back. I had been trying to draw attention to what people could now do with language models, and to the speed with which those capabilities were affecting intellectual work. A paper about something called AI could be historically relevant without settling whether its assumptions remained adequate. I wanted the change in the phenomenon to receive as much attention as the continuity of the subject’s name.
The AI revised its response explicitly:
“My earlier response gave the historical precedent too much weight and the change you are identifying too little.”
That concession mattered because it changed the next question. We could stop treating the discovery of a precedent as an answer to the possibility of methodological change. It did not, however, prove my stronger claim that an entirely new paradigm had arrived. We still had to distinguish historical novelty, technical change and the adequacy of an explanation.
My own response had also overstated the rupture. I spoke as though AI before 2022 and present language models might share little beyond a name. The technical chronology challenged that formulation: the transformer architecture appeared in 2017, and GPT-3’s account of large-scale language modelling in 2020. ChatGPT’s public introduction on 30 November 2022 belongs to that history (Vaswani et al., 2017); (Brown et al., 2020); (OpenAI, 2022). I could not use 2022 as a clean boundary between two unrelated technical worlds.
What survived was the claim about changes in intellectual practice. Considering an AI project as an object of theological reflection and repeatedly using a model to articulate and revise theology create distinguishable tasks. The same project can contain both. Technical continuity does not settle whether that combination changes the evidence, responsibilities and methods required. Nor did I need to claim that earlier thinkers had failed to imagine powerful AI. Anticipating a capability and working under the conditions created by its availability are different situations.
This left me with a more demanding position than simply announcing a paradigm shift. I would have to identify the assumptions that no longer hold, the practices that have changed, and the questions that an inherited approach leaves insufficiently examined. Alongside “which science?” and “which religion?”, I now needed to ask which AI system, performing which role, within which activity.
When I pushed back, I had also invoked Terence Tao’s “Math 2.0” as an indication that the issue extended beyond theology. The AI checked the reference and developed its relevance. Tao’s post of 10 October 2026 introduced slides from the previous evening’s Caltech lecture. He argues that mathematical institutions have been organised around the difficulty of obtaining proofs, and that increased automated problem-solving requires renewed attention to mathematical understanding and the health of the community. He also stresses uneven capabilities and the continuing difficulty of frontier problems (Tao, 2026).
The AI then proposed a connection to my earlier writing: producing a strong output and developing the person who produces it are distinguishable achievements. I had introduced Tao to emphasise the scale of technological change; the response made the purposes of a discipline more explicit. This connection needs a limit. Theological interpretation generally lacks the agreed formal checking available for appropriately formalised mathematical proofs. The analogy opens a question rather than supplying its answer: if fluent theological production becomes easier, what should count as evidence of theological understanding?
The relationship can enter the design of the work
Before that disagreement about historical precedent, my engineering background had prompted a blunt return to the reading list: had these approaches actually helped resolve anything important? I wanted examples in which the original authors did substantive work. The response required distinguishing what an author’s detailed argument achieved from what depended on the typology itself. Several cases made this distinction concrete.
Barbour’s discussion of Jacques Monod examines the passage from a biological account involving chance to the conclusion that there is no divine purpose. Barbour argues that “science does not deal with divine purpose,” rather than establishing its absence (Barbour, 1997, p. 80). Whatever one thinks of the wider theological position, the analytical task is clear: identify which conclusion is supported by the scientific account and which requires additional philosophical premises. Calling a position Conflict does not perform that analysis.
Stenmark pursues a more specific difficulty. If evolutionary history could have produced different organisms, how can a believer maintain that God intended human beings? He distinguishes a scientific account of contingency from the further premises needed to infer an absence of purpose. Then, provisionally allowing a stronger objection, he considers a theological revision: perhaps God’s intention concerns the emergence of free, self-aware beings rather than precisely the species Homo sapiens. His analogy is a parent intending to have a child without planning the exact individual who comes into existence (Stenmark, 2004, pp. 158–165).
That argument neither proves divine purpose nor settles what every tradition can accept. It does show a particular theological claim being examined and potentially revised in response to a particular challenge. Mivart’s case offered a different achievement: a person’s apparently inconsistent positions became more intelligible through their contexts. I had to allow that clarifying an inference or improving historical understanding could count as a real result. My engineering preference for an intervention was itself one criterion among others.
Chin’s hypothetical Christian biology student makes the choice of method equally concrete. A student wondering whether they can flourish in both communities may need historical examples or evidence about contemporary scientists’ experiences; an abstract demonstration of logical compatibility may leave the practical concern untouched. Chin presents this as guidance for choosing relevant scholarship, not a reported intervention whose educational success has been measured (Chin, 2023, pp. 673–674). I could now ask a more precise seminar question: what would we lose if we conducted this investigation without first assigning a relationship type?
For my own project, I still wanted to follow the consequences into practice. Could theological reflection change what a system was asked to do, which sources it treated as relevant, or what counted as an acceptable answer? Earlier writing on computational theology had led me to treat specification as theologically consequential: selecting a corpus, an authority structure and admissible reasoning shapes the inquiry before the system produces a conclusion (Yin, 2026b).
The present discussion changed the role of that earlier argument. It became a concrete way to test the usefulness of a science–religion framework. Calling the arrangement integration would locate a broad relationship. Understanding its consequences would require examining the actual choices through which the arrangement was constructed.
The AI suggested a possible investigation: hold a theological question constant while varying the texts available to the system. Does broadening a specified collection make disagreement visible, and does the answer accurately represent its sources and limits? Repeated trials and source comparison would be needed before attributing changes to the collection. This remained a proposed experiment, but it identified something more specific to investigate than whether computing and theology were integrated.
Technical precision matters here. Changing a source collection, altering retrieval, changing an instruction and redesigning a model’s architecture are different interventions. If I want to understand how theology influences computation, I need to identify where that influence enters. Otherwise, the word integration risks concealing the very mechanism I wanted to investigate.
A documented example also corrected one of my stronger claims. I had suggested that Claude’s constitution could be treated as a direct result of religious studies. Anthropic’s May 2026 account supports a narrower statement. It describes engagement with religious, philosophical and cultural communities. Within that wider effort, a discussion with scholars of neuroscience and character formation helped inspire an experiment in which Claude could call a tool that reminded it of its ethical commitments. The company reported reduced misaligned behaviour on several internal evaluations, while leaving unresolved how much improvement came from the reminder and how much from pausing (Anthropic, 2026).
This company report gives a specific route from reflection on moral formation to an engineering intervention. It does not establish an exclusively theological origin for the constitution, an independently validated general solution, or an equivalence between human virtue and model behaviour. Those distinctions make the case more useful. We can investigate a contribution without claiming that it settles the whole relationship.
Stenmark’s distinction among the problem-stating, development, justification and application phases of science becomes useful here. A religious commitment may influence what someone considers worth investigating or how a technology should be used. That influence differs from providing adequate justification for a claim about how well an intervention works (Stenmark, 2004, p. 267). The practical excursion therefore returned me to one of his distinctions with a clearer question: at which stage does the theological contribution enter, and what happens to it afterwards?
The same discipline applies to my own proposals. A theological account of humility might motivate attention to uncertainty or truthful reporting of limitations. Whether a particular intervention produces those behaviours remains a technical question. Whether the resulting behaviour adequately expresses what that tradition means by humility remains a theological question. The investigation may force revisions on both sides, without making their standards interchangeable.
Further reading during this revision also places my preference for specific projects within a current scholarly debate. Simon Maria Kopf’s discussion of science-engaged theology examines why, when and how scientific work becomes a source for theology. He supports specific engagements while challenging the idea that locality removes the need for methodological reflection or that this approach can replace the entire science-and-religion field (Kopf, 2026). This checks a temptation in my own argument. Beginning with a concrete AI task can make the inquiry more exact, but it does not decide which theological sources have authority or which technical findings warrant a conclusion. Moreover, using a computational tool to organise texts is distinguishable from treating scientific findings as a theological source. My project needs to specify which activity is occurring.
The inquiry changes its own investigator
Another continuity with my earlier work concerned the person doing the research. In Who Creates Whom?, I had distinguished the initiating self, the self who critically encounters an AI-assisted articulation, and the later self who returns to the preserved work. I called these positions you1, you2 and you3. They describe one continuing person at different moments of understanding and responsibility (Yin, 2026a).
One episode from that essay remained particularly relevant. An AI-assisted draft presented my conclusions coherently while compressing the process through which I had reached them. I changed the objective: the article needed to show how the ideas became thinkable. The interaction had altered my criterion for a successful result. I was bringing that earlier observation into a new methodological setting, rather than discovering the changing author for the first time.
In the present conversation, I wondered whether interactions among dimensions might generate new ones. The AI distinguished change within a dimension from the addition of a dimension. I did not find the distinction sufficient without another clarification: what exactly were we calling a dimension? Otherwise, we risked deciding whether my account fitted a vocabulary before explaining the vocabulary. The discussion then separated an analytical aspect of inquiry from an independent coordinate in a mathematical model. The three successive authorial positions do not automatically establish three mathematical dimensions.
The AI supplied a hypothetical example that clarified what an additional descriptive variable might achieve. Two researchers could hold the same present beliefs while having different histories: one had published a misleading claim and might owe a correction; the other had not. A description containing only their present beliefs would miss this difference in responsibility. That made the question more useful: new relative to which existing account, and necessary to explain what? The point was neither that every change creates a dimension nor that an existing vocabulary must always be sufficient.
My question about cybernetics also brought us to Ashby’s distinction between the thing being investigated and the variables selected to represent it. His account of a system makes the choice of variables part of the work of inquiry (Ashby, 1956, section 3/11). This helped me take my own demand for specificity seriously. Calling the process open, dynamic or reflexive could orient attention, but I still had to say what changed and why it mattered.
Alongside this dispute about dimensions, I was trying to describe the process itself. I pictured an inquiry receiving something from outside, producing a response, and then encountering that response as material for another change. I asked whether cybernetics or a related language could help. The AI had already proposed coupled update relations in response to my earlier question about a systematic or mathematical account. I now asked it to explain what they actually meant:
\[
\begin{aligned}
x_{t+1} &= F(x_t,c_t,u_t),\\
c_{t+1} &= G(c_t,x_t,u_t).
\end{aligned}
\]
Here, \(x_t\) represents the developing work, including the relevant recorded history; \(c_t\) represents the criteria used to judge it; and \(u_t\) represents a new input, such as a reading, objection or AI response. The first relation says that the work changes through the encounter. The second allows the encounter to change the standards by which the next version will be judged. Their coupling expresses the possibility that changes in the work and changes in judgment affect one another.
These equations have a deliberately limited status. The AI suggested them as a representation of the process we were discussing. We did not specify or estimate the functions \(F\) and \(G\), demonstrate predictive power, or establish a new mathematical theory of learning. The notation helped isolate a question. It should face the same demand I made of the typologies: what does this description enable us to understand, and what evidence would justify using it?
The movement outward and back can be described with this distinction in mind. At first, an adequate seminar answer seemed to mean an accurate comparison of the frameworks. After the discussion of purposes, it needed to identify what the comparison was for. After the objections about asymmetry, it needed to establish whose problem was being addressed. After the demand for concrete achievements, it needed to distinguish the usefulness of an argument from the usefulness of its classification. The subject remained the readings; the standards I brought to them became more explicit.
Looking for scholarship to interpret this pattern during revision, I returned to Donald Schön’s work on design. His account describes a practitioner making a move, encountering its consequences, and seeing the situation differently. He also warns against “reading back onto the beginning of a process what has emerged only at its end” (Schön, 1992, pp. 132–133). The relevance is methodological: a later account can make the problem appear fully defined before the inquiry helped define it. My conversation was not an architectural design exercise, but his analysis helps explain why a polished comparison table could omit the intellectual work I most wanted to preserve.
Akkerman and Bakker’s review of boundary crossing offers another useful distinction. Learning across practices can involve articulating one’s own perspective and coming to understand another; deeper transformation involves changes in practices themselves (Akkerman and Bakker, 2011, pp. 144–147). This gives me a way to examine what happened when an engineering expectation of practical results encountered historical or theological standards of achievement. It does not mean that every change of topic constitutes boundary crossing, or that my preparation demonstrated a transformed discipline. The evidence here is more modest: some encounters changed how I understood and evaluated the task I returned to.
There is an evidential danger here. An AI can refine a criticism in the direction its user prefers. A later, more agreeable formulation is not independent confirmation of the user’s position. My repeated resistance could help uncover a neglected distinction, or it could encourage a persuasive rationalisation of my initial dissatisfaction. The original texts mattered because they supplied constraints that neither conversational agreement nor the elegance of the emerging story could replace.
The archive creates a related difficulty. It allows me to inspect earlier formulations, rejected suggestions and changes of direction. Yet this article is itself a later reconstruction, written with AI assistance, selecting which moments count as consequential. Preserving a conversation does not remove interpretation from the account of it. I remain responsible for making that selection answerable to the record, including the moments in which my criticism needed correction.
The successive drafts make the difficulty concrete again. I first asked that the history of the thinking be restored to a polished account. Reading the next version, I could see more of the objections and corrections, yet something still seemed missing: I had repeatedly widened the inquiry and then deliberately returned to the seminar task. A draft could preserve individual changes of mind while losing that organising movement. The present revision responds to that further judgment. Its structure must show how a broader question changed the next reading, and how the reading constrained the broader claim.
Returning to the seminar with a different question
I can now give a more discriminating response to the original assignment. Typologies can orient a reader, make alternatives visible and support comparisons. Their weakness emerges when those achievements are mistaken for an explanation of how an encounter develops. For tracing the changing purposes, standards and organisation of an AI-assisted inquiry, Stenmark gives me the most useful starting vocabulary among the relationship frameworks considered here. Barbour’s four broad categories, used alone, give me the least detail for that particular task, while remaining useful for an introduction. Loke improves precision about what is related and whose perception is being classified. Chin helps me choose an investigative method; evaluating him as though he offered another first-order relationship taxonomy would miss his purpose.
The alternatives therefore remain integral to my answer. Biography asks how commitments operate within a life; sociology investigates practices and institutions; conceptual analysis tests inferences; attention to language examines how a question is framed. Gregersen’s account of a field with several disciplinary arms and a shared centre offers a helpful reminder that expansion needs an organising concern (Gregersen, 2014, pp. 419–420, 425–428). His metaphor describes an academic field, while I am describing a process of preparation. The connection I draw is that reaching into another discipline becomes useful when I can explain what it brings back to the problem.
My alternative would begin with a specified activity and follow its development. In AI-assisted theological writing, I would preserve the originating question, the relevant sources, substantial AI contributions, decisive objections and subsequent revisions. I would ask what changed in the argument and what changed in its evaluation. Later rereading would matter where it produced a correction, a further question or a demonstrable change in practice. Recording the sequence would provide material for analysis; its existence alone would not prove that learning had occurred.
This approach draws on resources already present in the readings: attention to lives and contexts, distinctions among dimensions of practice, and scrutiny of scholarly methods. Its possible contribution lies in investigating particular contemporary mechanisms. How does a generated interpretation alter the question a theologian asks next? When does changing a source collection expose a hidden assumption? Under what conditions does repeated assistance strengthen a researcher’s judgment, and when does it merely make an unstable interpretation easier to express?
I would provisionally call my way of working a recursive, problem-centred inquiry. This describes the movement I can identify in the record; it is not a claim to have invented a new educational method. I begin with an obstacle in understanding, pursue a wider connection far enough to test it, and return to the focal task with a revised distinction, objection or criterion. The return matters as much as the expansion. A connection earns its place when it changes what I can ask of the evidence or what I can responsibly conclude.
This also gives the process a practical limit. I cannot follow every association indefinitely. Sometimes a source corrects the premise of an excursion; sometimes the discussion has moved to another question and needs to be redirected; sometimes I have enough to formulate a provisional seminar contribution while leaving the larger investigation open. Adaptability includes deciding when to widen the inquiry, when to return, and when to retain a question without pretending to have resolved it.
I began with the feeling that AI and theology worked together so naturally that their relationship scarcely needed explanation. I now think that this apparent ease is one reason to examine the relationship closely. The interaction can shape which questions become available, how answers are judged, and what I learn to regard as adequate theological work. The question I want to take back to the readings is therefore: how well can our frameworks investigate an encounter that helps change the practices, purposes and people through which the encounter itself is understood?
References
Akkerman, S. F., and Bakker, A. (2011). “Boundary Crossing and Boundary Objects.” Review of Educational Research, 81(2), 132–169. https://doi.org/10.3102/0034654311404435.
Anthropic (2026). “Widening the conversation on frontier AI.” 19 May.
Ashby, W. R. (1956). An Introduction to Cybernetics. London: Chapman & Hall. Author’s estate-authorised digital edition.
Barbour, I. G. (1997). Religion and Science: Historical and Contemporary Issues. Revised and expanded edition. San Francisco: HarperSanFrancisco.
Barbour, I. G. (2002). “On Typologies for Relating Science and Religion.” Zygon, 37(2), 345–360. https://doi.org/10.1111/0591-2385.00432.
Brown, T. B., et al. (2020). “Language Models are Few-Shot Learners.” Advances in Neural Information Processing Systems, 33. https://arxiv.org/abs/2005.14165.
Cantor, G., and Kenny, C. (2001). “Barbour’s Fourfold Way: Problems with His Taxonomy of Science–Religion Relationships.” Zygon, 36(4), 765–781. https://doi.org/10.1111/0591-2385.00395.
Chin, A. J. (2023). “The Aims of Typologies and a Typology of Methods.” Zygon, 58(3), 656–677. https://doi.org/10.1111/zygo.12890.
Drees, W. B. (2010). Religion and Science in Context: A Guide to the Debates. Abingdon: Routledge.
Evans, J. H., and Evans, M. S. (2008). “Religion and Science: Beyond the Epistemological Conflict Narrative.” Annual Review of Sociology, 34, 87–105. https://doi.org/10.1146/annurev.soc.34.040507.134702.
Foerst, A. (1998). “Cog, a Humanoid Robot, and the Question of the Image of God.” Zygon, 33(1), 91–111. https://doi.org/10.1111/0591-2385.1291998129.
Gregersen, N. H. (2014). “Prospects for the Field of Science and Religion: An Octopus View.” Zygon, 49(2), 419–429. https://doi.org/10.1111/zygo.12091.
Kopf, S. M. (2026). “Prospects and Pitfalls of Science-Engaged Theology.” Religious Studies, 62(2), 305–328. First published online 2 May 2025. https://doi.org/10.1017/S0034412525000307.
Lee, A. (2022). “Seeing the World Through the Languages of Science and Religion.” In B. Sollereder and A. McGrath (eds), Emerging Voices in Science and Theology: Contributions by Young Women, pp. 28–42. Abingdon: Routledge. https://doi.org/10.4324/9781003251446-4.
Loke, A. (2023). “A New Fourfold Taxonomy of Science-Religion Relations.” Theology and Science, 21(1), 29–43. https://doi.org/10.1080/14746700.2022.2155909.
OpenAI (2022). “Introducing ChatGPT.” 30 November.
Schön, D. A. (1992). “Designing as Reflective Conversation with the Materials of a Design Situation.” Research in Engineering Design, 3, 131–147. https://doi.org/10.1007/BF01580516.
Stenmark, M. (2004). How to Relate Science and Religion: A Multidimensional Model. Grand Rapids, MI: William B. Eerdmans.
Tao, T. (2026). “Math 2.0.” What’s New, 10 October. Includes slides from the Caltech lecture delivered on 9 October.
Vaswani, A., et al. (2017). “Attention Is All You Need.” Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762.
Yin, R. (2026a). “Who Creates Whom? Art, AI, and the Transformation of the Maker.” 18 August.
Yin, R. (2026b). “Toward a Boundary-Relative Computational Theology: What Must Theology Become to Be Verifiable?” 21 August.