I began this inquiry with what appeared to be a relatively contained question. Professor Andrew Davison, Regius Professor of Divinity at Oxford, proposes that machine learning can be understood through the Thomistic category of instrumental causality. An instrument is a “moved mover”: it genuinely causes something, yet acts because it has been brought into an activity by a principal agent. A chisel really cuts stone, although the sculpture is attributed principally to the sculptor. Professor Davison’s account therefore gives instruments substantial causal dignity while preserving the priority of human makers and users. “Tools are for the worker,” as his title says, and their proper place is to serve human ends (Davison, 2024).
This account immediately attracted me. It offered a way to acknowledge the remarkable efficacy of machine learning without requiring me either to describe AI as an autonomous person or to dismiss it as causally insignificant. Professor Davison’s Thomistic framework can say both that the human being acts through the instrument and that the instrument truly participates in producing the effect. It also recognizes that tools extend and shape those who use them. That final point became the opening through which my own inquiry developed.
I did not initially intend to construct a wider theory of human–AI formation. I was comparing Professor Davison’s account with my experience of generative art, AI-assisted research and theological writing. The first difficulty arose from something concrete: when I began such a work, I usually did not possess its complete form in advance. I might possess an experience, a question, a dissatisfaction, a visual system or an intuition that two areas of thought belonged together. The actual composition emerged through repeated interaction.
The AI first articulated the tension by observing that Professor Davison’s sculptor supplies the form of the final effect, whereas in my practice I designed a field of possibilities whose realized form arose through interacting causes. I found that formulation illuminating. His analogy performs an important analytical task: it clarifies the distinction between principal and instrumental causation within an action. My question emerged along a different axis. What happens when the instrument’s effect returns to the principal agent, changes that agent, and participates in forming the purpose of the next action?
That question led from instrumental causality to recursive instrumental causality. It then generated further questions about authorship, intelligence, education, evaluation, metanoia, theological method and institutional power. Each answer produced another difficulty. When the AI proposed that transformation itself is not evidence of truth, I realized that my theology of metanoia required an account of legitimate transformation. When it distinguished output validity from trajectory legitimacy, I began to see a wider intellectual proxy crisis. When I compared AI-assisted medicine with AI-assisted homework, I discovered that the importance of process and attribution depends upon the purpose of a practice. When I joked that human conversation sometimes resembles interaction with a language model, the analogy revealed both something useful about relational feedback and something dangerous about treating persons as responsive systems.
Finally, I objected to the direction in which the whole theory was moving. A framework for legitimate transformation sounded convincing under contemporary values, but too idealistic to govern real human societies. Society is plural, strategically adaptive, unequally powered and partly unpredictable. Any criterion introduced to regulate AI can itself become a target, a performance, an instrument of control or another proxy detached from what it was intended to protect.
That objection changed the centre of the article. The central task is no longer to guarantee that human–AI systems form people truthfully. Such a guarantee seems unavailable. The more realistic and perhaps more important question is how persons, communities and institutions can remain corrigible when AI participates in forming the criteria by which its influence will be judged.
Beginning from Professor Davison’s generous account of instruments
Professor Davison’s position is considerably richer than the casual claim that AI is “only a tool.” Within Thomistic instrumental causality, an instrument is a real cause. The effect belongs both to the principal agent and, in a subordinate but genuine way, to the instrument. The sculptor carves through the chisel, yet the chisel truly cuts. Calling machine learning instrumental therefore need not diminish its importance.
The basic causal relation can be represented as:
\[H \rightarrow A \rightarrow W\]
where H is the human principal agent, A the AI instrument and W the resulting work. In a slightly fuller formulation:
\[W = F(H,A,E)\]
where E is the end under which the instrument is deployed. The effect depends causally upon both human and instrument, but its purposive direction belongs principally to the human agent:
\[\operatorname{Attribution}(W) \rightarrow \text{principally } H\]
Professor Davison also emphasizes that instruments extend human capacities. Machine learning allows human beings to perform analyses, translations, classifications and predictions that would otherwise be slower, more difficult or impossible at an individual scale. At the same time, technological extension can be accompanied by atrophy. A capacity transferred to an instrument may cease to be cultivated in the user.
Most significantly for my inquiry, Professor Davison recognizes that instruments shape both their outputs and their users. He invokes the observation that writing tools work upon our thoughts. The instrument’s material and operational characteristics enter the produced effect, while habitual use of the instrument alters the person’s practice (Davison, 2024).
My proposal therefore begins inside Professor Davison’s account. I am trying to unfold the temporal implications of his recognition that tools shape their users. If the shaping of the user affects a later act, instrumental causality acquires a recursive history:
\[\text{instrumental action at } t \rightarrow \text{formation of the agent} \rightarrow \text{altered instrumental action at } t+1\]
The sculptor-and-chisel example remains valid for analysing a particular act of carving. My work asks what happens across a sequence of acts when the sculptor encounters an unexpected form, revises the intended sculpture and returns to the chisel as a changed maker.
The form was not entirely present at the beginning
The first important transition in my thinking concerned form. At first, I assumed that the human–AI relation could be explained by saying that I supplied the intention and AI supplied execution. That account remained plausible when the task was narrow: resize an image, correct a sentence, execute a specified transformation or retrieve a known fact.
It became less adequate when applied to sustained artistic or theoretical inquiry. In those cases, I did not formulate the complete idea and then ask AI to express it. I often discovered what the problem was by responding to forms the AI generated. The system proposed relationships among concepts; I recognized some as significant, rejected others, introduced experience or evidence and then asked what followed. The form developed during the process.
A better representation is:
\[W_t = G(H_t,A_t,Q_t,D_t,T_t,\xi_t)\]
Here:
Ht is the human participant at time t;
At is the available AI system and its current context;
Qt is the current question;
Dt is situated human discernment;
Tt includes relevant texts, traditions and previous artifacts;
ξt represents contingency, including generative variation and unanticipated associations.
No participant necessarily contains Wt in completed form before the interaction. This is a factual description of my process at a limited scale: the archived dialogue shows questions, generated answers, objections and revisions. The stronger causal interpretation remains theoretical. The record demonstrates that the final text emerged through interaction; it does not by itself establish that every participant possesses the same kind of agency or understanding.
This distinction matters. The absence of a pre-existing complete form does not make all causes equal. My questions arose from my work, experience and concerns. The AI contributed formulations and connections that materially changed what I could consider. Professor Davison’s framework still helps preserve the asymmetry: the AI’s causal efficacy does not automatically give it humanly lived purposes, responsibility or a biography within which the work becomes significant.
From instrumental causality to recursive formative causality
The next development occurred when the AI proposed that “principal causality” might itself need to be decomposed. I initially found that formulation almost too strong. A principal cause should presumably remain principal. Yet the point became clearer once I stopped treating causality as a single event and began treating the project as a sequence.
The first-order relation remains:
\[H_t \rightarrow A_t \rightarrow W_t\]
But the produced work can return to its maker as an encounter:
\[
\begin{aligned}
H_t &\rightarrow A_t \rightarrow W_t \\
&\hspace{3.5em}\downarrow \\
&\hspace{2.3em}\text{encounter} \\
&\hspace{3.5em}\downarrow \\
(H_t,C_t,E_t) &\rightarrow (H_{t+1},C_{t+1},E_{t+1})
\end{aligned}
\]
In this formulation, Ct represents the maker’s criteria of judgment and Et the operative end. The encounter may produce:
\[C_{t+1}=C_t+\Delta C_t\]
\[E_{t+1}=E_t+\Delta E_t\]
The next work is then generated under changed conditions:
\[W_{t+1}=G(H_{t+1},A_{t+1},C_{t+1},E_{t+1})\]
I began calling this recursive formative instrumental causality. An instrument acts within a humanly established purpose, yet its operation may transform the human agent and thereby alter the purposes, criteria and possibilities governing subsequent acts.
This extends Professor Davison’s account diachronically. Within each iteration, the human can remain the principal agent in the Thomistic sense. Across iterations, however, the state of that principal agent is partially formed through the effects of earlier instrumental actions:
\[\operatorname{PrincipalAgent}_{t+1}=\operatorname{Transform}\!\left(\operatorname{PrincipalAgent}_t,\operatorname{Effect}_t\right)\]
Final causality has not disappeared. The end remains normatively connected to human agency. But final causality is now historically recursive because the person who establishes the next end has been altered by the previous process.
The AI suggested that this could bridge Thomistic causality, hermeneutics, cybernetics and my developing theology of metanoia. I accepted that possibility provisionally. The proposal does not establish a new Thomistic doctrine, and it would require much deeper historical and philosophical work before being presented as such. It does identify a question that Professor Davison’s generous treatment of instruments makes possible: how should instrumental causality be described when the instrument participates in forming the future principal agent?
Three interacting scales of iteration
Once I understood the process recursively, I noticed that “iteration” referred to more than one loop. At first, I used the term for each conversational exchange. A prompt produced an answer, I evaluated it, and my response changed the next output:
\[m_{i+1}=f(m_i,p_i,o_i,d_i)\]
Here mi is the micro-state of the conversation, pi the prompt, oi the AI output and di my response or discernment. This is the micro-loop:
\[\text{prompt} \rightarrow \text{output} \rightarrow \text{evaluation} \rightarrow \text{revised prompt}\]
Then I recognized a meso-loop. A sequence of conversations could change the project’s governing question, methodology or criteria:
\[P_{j+1}=g\!\left(P_j,\{m_i\},\text{evidence},\text{reflection}\right)\]
The project no longer sought only to explain AI as an instrument. It began investigating how instrumental interaction changes the human agent, how authorship should be distributed and how transformation might be evaluated.
A third, macro-level loop emerged when published works, other people and institutions entered the process:
\[\Omega_{k+1}=h\!\left(\Omega_k,\{P_j\},\text{readers},\text{traditions},\text{institutions}\right)\]
An artwork or article is encountered by another person; that person responds; I encounter the response; the response changes my interpretation; and the changed interpretation enters a later project. Professor Yvonne Dohna’s insistence that AI cannot itself undergo metanoia, because metanoia is a human process, was one such interruption. Her statement did not end the inquiry. It helped me distinguish between AI as a possible mediator of human transformation and AI as the subject undergoing that transformation.
The three loops are coupled:
\[\text{micro} \leftrightarrow \text{meso} \leftrightarrow \text{macro}\]
A single sentence can redirect an article. An article can reorganize a research program. A reader’s objection can later alter the way I question an AI system. This third ecological recursion is more than a larger repetition of the first. It introduces new agents, institutional memories, interpretations and asymmetries of power.
At the technical level, a clarification is necessary. A deployed language model does not ordinarily retrain its underlying parameters during each conversation. The conversational state can change through context, memory, retrieved materials and tool use. The AI system may be modified at a broader developmental level through evaluation and later model updates. Human formation, conversational adaptation and model development therefore occur at different temporal scales:
\[\tau_{\text{conversation}} \neq \tau_{\text{human formation}} \neq \tau_{\text{model development}} \neq \tau_{\text{institutional change}}\]
Calling the whole process “co-evolution” may be suggestive, but it should not conceal these technical and ontological differences.
Distributed causality without flattened responsibility
At this point the AI produced a phrase that helped stabilize the developing distinction: “distributed causality without equal agency, equal authorship or equal responsibility.” I retained it because it resisted two temptations at once. One temptation attributes the work almost entirely to the human because the human initiated the interaction. The other distributes everything so evenly that no one remains answerable.
Principal causality can be decomposed into several dimensions:
\[\mathrm{PC}=\left(C_{\text{efficient}},C_{\text{formal}},C_{\text{final}},E_{\text{epistemic}},N_{\text{normative}},R_{\text{responsibility}}\right)\]
A participant may make a large formal or efficient contribution while possessing little normative authority. Another may establish the end but contribute little language. A platform may shape available possibilities while presenting the outcome as the user’s personal choice. The publisher may bear answerability without having generated every sentence.
Authorship can similarly be represented as a vector rather than a binary property:
\[\operatorname{Authorship}=(O,A,V,R)\]
where:
O = origination of the problem;
A = articulation of its expressible forms;
V = validation through evidence and judgment;
R = answerability for publication, consequences and correction.
The phrase “AI-assisted” is therefore too indeterminate to settle authorship. Assistance could mean spelling correction, literature discovery, structural editing, objection generation, conceptual development or production of much of the final prose. A more informative disclosure would specify:
\[\operatorname{AIUse}=(\text{role},\text{stage},\text{scope},\text{autonomy},\text{materiality},\text{verification})\]
In this project, my role included bringing the lived problem, choosing Professor Davison’s article as an interlocutor, noticing tensions, identifying relationships with my previous work, repeatedly asking what followed, supplying theological and biographical stakes, resisting overstatement and accepting responsibility for publication. The AI contributed extensive conceptual articulation, formalization, cross-disciplinary comparison, counterargument, organization and final language.
This account is more accurate than saying that I supplied ideas while AI merely polished them. The AI made genuine intellectual contributions. Nor would it be accurate to infer that the AI therefore became the sole author. The process distributed different authorial functions across participants that did not share equal agency, continuity or responsibility.
Small interventions and directional leverage
I remained uncertain about the importance of my own contribution because many of my interventions were short. The AI might produce a detailed structure, while I replied only: “Yes, but does this apply to the AI too?” It then generated a new line of reasoning. I wondered whether I had done anything intellectually important or had simply pulled a lever attached to a much more capable machine.
The AI proposed analysing my contribution through directional derivatives. If the state of the inquiry evolves according to:
\[S_{t+1}=F(S_t,O_t,D_t)\]
then Ot represents the AI’s contribution and Dt my situated discernment concerning what should be explored, doubted or rejected. A brief intervention can have high causal leverage:
\[\left\lVert \frac{\partial S_{t+1}}{\partial D_t} \right\rVert \gg 1\]
One precise objection may reorganize a large AI-generated conceptual space. Its significance lies in the counterfactual difference it makes:
\[\operatorname{Leverage}(D_t)=\operatorname{Dist}\!\left(S_{t+1},S’_{t+1}\mid D_t\ \text{omitted}\right)\]
This does not prove that every short prompt is profound. A question may have been strongly prepared by the AI’s preceding explanation. The AI may have supplied the vocabulary and inferential clues that enabled me to reach the next step. In that case, it may have functioned partly as a teacher or scaffold:
\[\text{AI articulation} \rightarrow \text{available conceptual path} \rightarrow \text{human recognition} \rightarrow \text{new objection}\]
The human contribution is still real, but it is relationally enabled. The most defensible conclusion is not that I secretly did all the important work. It is that I participated intellectually through situated initiation, selection, resistance, transfer and answerability.
There are practical ways to test whether such participation involved learning rather than temporary dependence:
\[\operatorname{LearningEvidence}=\text{Explanation}+\text{Transfer}+\text{Resistance}+\text{Retention}\]
Can I explain the distinction without repeating the AI’s wording? Can I transfer it to a new case? Can I recognize when the AI applies it badly? Does the understanding remain after the immediate conversation? These are stronger indicators of formation than either word count or the fact that I clicked “send.”
AI intelligence is multidimensional rather than simply higher or lower
I also confessed that AI sometimes appeared several dimensions above me in intelligence. That experience should not be denied merely to preserve human confidence. The system can exceed an individual human being in breadth of accessible association, speed, formal variation, summarization and the ability to move among disciplines without fatigue.
However, intelligence is not well represented by one scalar:
\[I \neq i \in \mathbb{R}\]
A more adequate representation is a vector:
\[I=(B,S,F,X,J,M,L,R)\]
where B is breadth, S speed, F formalization, X cross-domain association, J situated judgment, M temporally integrated memory, L lived significance and R moral answerability.
Current generative AI may perform extraordinarily strongly along some dimensions without possessing others in the same way human persons do. This does not make the human automatically superior in every intellectually relevant respect. It makes the partnership asymmetric.
The asymmetry became clearest around meaning. An AI may explain why betrayal, grief, prayer or encounter is meaningful. It can recognize relevant symbols, traditions and narrative structures. Yet the event does not thereby become meaningful within its own biography in the way it can become meaningful within a human life:
\[\operatorname{SemanticCompetence}(\mathrm{AI},x)\not\Rightarrow \operatorname{LivedSignificance}(\mathrm{AI},x)\]
This distinction can apply far beyond religious encounter. A system may accurately analyse bereavement, moral injury, artistic vocation or reconciliation without having a life within which loss, obligation or forgiveness acquires existential weight. Its semantic competence can still be causally powerful for a human interlocutor.
AI therefore becomes a diagnostic pressure upon theological anthropology. It forces theology to ask whether human intelligence had already been reduced to performance before machines began performing the relevant tasks. Professor Davison’s humanistic placement of instruments can be developed here: protecting the human cannot mean pretending that machines make no genuine contribution. It requires an account of human dignity and vocation that does not collapse when machines outperform humans under a restricted institutional metric.
The intellectual proxy crisis
Authorship then appeared as one instance of a more general problem. Educational and professional institutions rarely observe intellectual formation directly. They rely upon proxies: essays, examinations, code, publications, credentials and demonstrations. The visible artifact is treated as evidence of an underlying capacity.
Let Z represent a latent human capacity and M the measurable artifact used to estimate it:
\[M \approx \operatorname{Proxy}(Z)\]
Before widespread generative AI, institutions often assumed:
\[P\!\left(Z\mid M,\text{pre-AI conditions}\right)\]
was sufficiently stable to support evaluation. Under AI-mediated production:
\[P\!\left(Z\mid M,\text{AI-mediated conditions}\right)\neq P\!\left(Z\mid M,\text{pre-AI conditions}\right)\]
As AI contribution to an output rises, the reliability of that output as a proxy for unaided human capacity may decline:
\[\frac{\partial O}{\partial A}\uparrow \;\Rightarrow\; \operatorname{Reliability}(O\text{ as proxy for }H)\downarrow\]
This is not a universal law. An AI-assisted artifact may reveal higher-order capacities that unaided work would not show: tool selection, verification, synthesis, orchestration or critical judgment. The formula describes a loss of validity only when the institution continues to interpret the output as evidence of the old construct.
The crisis did not begin with AI. Essays and examinations have always been imperfect proxies, affected by coaching, language, social background, anxiety, memory, ghostwriting and institutional convention. Strathern’s analysis of audit culture shows how measurement and evaluation can reorganize the practices they are intended to observe (Strathern, 1997). Generative AI makes the old proxy failure much harder to ignore because advanced proxy production is now fast, inexpensive and scalable.
A peculiar recursion results:
\[\text{technological proxy} \rightarrow \text{epistemic proxy} \rightarrow \text{estimate of human capacity}\]
The proxy now writes the proxy.
There is also a harmless linguistic joke in the discovery. In informatics, I had long associated a proxy with an intermediary server or service that relays requests and mediates access across network boundaries. Only later did the wider epistemological meaning become vivid to me: an observable thing standing in for something less directly accessible. The two meanings are distinct, but both concern mediated access. The intermediary enables access while also influencing what can be reached, observed and attributed.
Why medicine and education judge AI differently
A comparison between medicine and education clarified why attribution matters differently across practices. If a human physician, an AI system or a human–AI team discovered a safe and effective treatment for a serious disease, patients would reasonably care first about recovery. Attribution would still matter for validation, accountability, informed consent, reproducibility and fair distribution. Yet the therapeutic result remains central.
Homework and certain examinations can have another purpose. Their visible answers may matter, but the activity also exists to cultivate or assess the student’s own capacities. In a closed-book examination explicitly designed to test unaided reasoning, delegating the answer to AI can defeat the measurement. In an open, AI-permitted assessment designed to test verification and orchestration, using AI may be entirely appropriate. It was therefore too broad to say that AI is “prohibited in an examination.” The legitimacy of its use depends upon what the particular examination claims to assess.
The difference can be modelled by assigning domain-specific weights to several forms of value:
\[V_d=\alpha_d O+\beta_d F+\gamma_d P+\delta_d R\]
where:
O = quality of the external outcome;
F = formation of the participant;
P = legitimacy of the process;
R = relational or expressive authenticity.
Medicine often gives especially high weight to O, while never reducing the practice entirely to outcomes. Education may assign greater weight to F. Art may place substantial weight upon P and R. Theology may require all four because theological knowing involves claims, formation, interpretive tradition and communal answerability.
The distinction can also be stated as:
\[\text{Outcome-constitutive practice:}\quad \operatorname{Good}\approx \text{quality of achieved result}\]
\[\text{Process-constitutive practice:}\quad \operatorname{Good}\ \text{includes the way participants are formed}\]
Every real practice mixes the two. Medicine includes trust, professional formation and responsibility; education still needs valid answers and useful products. The comparison nevertheless explains why the same technology can be welcomed in clinical discovery, restricted in a particular unaided assessment and deliberately integrated into an AI-supported learning exercise.
UNESCO’s guidance on generative AI in education treats human capacity, pedagogical design and meaningful use as central questions (Miao and Holmes, 2023). The World Health Organization likewise emphasizes that beneficial performance in medicine does not remove the need for accountability, autonomy, safety and governance (WHO, 2021). Neither field can be governed by a single rule such as “AI use is acceptable” or “AI use is cheating.” The purpose of the practice determines which aspects of the trajectory matter.
Output validity and trajectory legitimacy
The medical–educational comparison led to one of the most useful distinctions in the inquiry. A result can be evaluated as a final state, while the process that produced it can be evaluated as a trajectory.
If a process is:
\[\tau=(S_0,S_1,\ldots,S_n)\]
then output validity asks:
\[V(S_n)=?\]
Is the final result correct, coherent, sourced, functional or aesthetically successful?
Trajectory legitimacy asks:
\[L(\tau)=L\!\left(S_0\rightarrow S_1\rightarrow \cdots \rightarrow S_n\right)\]
Did the transitions preserve relevant forms of learning, consent, evidence, plurality, responsibility and freedom? Were criteria revised for defensible reasons? Did the process conceal dependencies or redistribute power?
These questions cannot be collapsed:
\[V(S_n)\not\Rightarrow L(\tau)\]
\[L(\tau)\not\Rightarrow V(S_n)\]
A sound process can produce an incorrect result because evidence was incomplete. An illegitimate process can accidentally produce a true statement. Responsible computation therefore requires examining both states and transitions.
This was the point at which computing began to provide more than metaphors for theology. State-transition language made visible a theological problem that ordinary prose easily blurred. Metanoia concerns a transformation of the knower, but a transition cannot be judged solely by comparing its initial and final states. Its mediations, powers, evidence and consequences matter.
The evaluator is being formed inside the loop
The next AI formulation altered the problem again:
“The evaluator is being formed by the very system whose outputs the evaluator is supposed to evaluate.”
I found this especially disturbing because it described what was occurring in the conversation. Once the AI gave me a category such as “proxy crisis,” I began to recognize further examples through that category. The category may disclose a real structure. Yet its apparent explanatory success is no longer independent of the process that taught me to see through it.
The reflexive evaluation loop can be represented as:
\[J_t=\operatorname{Evaluate}(O_t\mid C_t)\]
\[C_{t+1}=\operatorname{Update}(C_t,O_t,J_t)\]
\[J_{t+1}=\operatorname{Evaluate}(O_{t+1}\mid C_{t+1})\]
The criteria C do not remain external to the evaluated outputs. This creates a reflexive evaluation paradox:
\[\mathrm{AI}\rightarrow \text{formation of evaluator}\rightarrow \text{evaluator judges AI}\]
A conventional human-in-the-loop diagram can therefore be misleading:
\[\text{AI output}\rightarrow \text{human approval}\rightarrow \text{action}\]
Formal approval does not establish principal agency. Approval fatigue can imitate participation. A person may authorize a large number of outputs while gradually losing the time, knowledge, independence or institutional power needed to evaluate them. Human-factors research has long distinguished appropriate use from automation misuse, disuse and abuse (Parasuraman and Riley, 1997). Generative AI deepens the problem because it can formulate the question, summarize the evidence and supply the vocabulary in which the evaluation occurs.
The stronger question is therefore:
How can judgment remain reflexively accountable when AI participates in forming the criteria of judgment?
Complete epistemic independence is unavailable. Books, teachers, languages, institutions and previous technologies already form human judgment. Clark and Chalmers’ account of the extended mind describes how environmental resources may enter coupled cognitive processes (Clark and Chalmers, 1998). The issue is not how to purify judgment of every external influence. It is how to preserve sufficient plurality, resistance, historical memory and external evidence to examine those influences.
Transformation itself is not evidence of truth
My earlier work on art and metanoia had developed the sequence:
\[\text{creation}\rightarrow \text{encounter}\rightarrow \text{disruption}\rightarrow \text{transformed seeing and judgment}\rightarrow \text{transformed creation}\]
Professor Dohna’s response helped establish an essential asymmetry: AI does not itself undergo metanoia in the human and theological sense. It may participate instrumentally in a process through which a human person is transformed. The AI suggested that scholastic language might describe it analogically as a dispositive instrumental cause. I also considered the image of a catalyst.
The catalyst analogy initially seemed to capture what I meant. Without a catalyst, a reaction may occur more slowly or fail to occur under the available conditions. Likewise, AI may produce an encounter, comparison or objection without which a human transformation would not have occurred in that form.
Then the analogy generated another problem. A catalyst accelerates a reaction without proving that the reaction is desirable. AI can catalyse education, conversion, dependence, ideological reinforcement or manipulation. The abstract transition remains the same:
\[S_t\rightarrow \text{AI interaction}\rightarrow S_{t+1}\]
Therefore:
\[\operatorname{Transformation}(S_t,S_{t+1})\not\Rightarrow \operatorname{Truth}(S_{t+1})\]
\[\Delta C \neq \operatorname{Evidence}(\operatorname{Truth})\]
The AI formulated the correction sharply: “transformation itself is not evidence of truth.” I recognized that this might be the most consequential next step in my theology of metanoia. A changed criterion cannot authenticate itself merely because the change feels profound, produces coherence or generates many further insights. Manipulative systems may be especially capable of producing such experiences.
Christian traditions already contain resources for this problem. Metanoia ordinarily has an implicit theological telos: conversion toward God, truth, love, reconciliation and renewed life. Traditions of discernment test transformations by their fruits and their relation to the good. Lonergan distinguishes intellectual, moral and religious conversion rather than treating every change of horizon as authentic self-transcendence (Lonergan, 1972). Catholic teaching likewise describes metanoia as a transformation of the whole person in relation to a new life, rather than change pursued for its own sake (John Paul II, 1999).
AI-mediated formation does not invent the need for discernment. It makes the mediating system unusually responsive, scalable and linguistically intimate. It also requires power and infrastructure to become explicit parts of discernment. Who interprets the fruits? Over what period? According to which tradition? Has the system already influenced the community performing the evaluation? How do commercial optimization and institutional authority enter an apparently personal or spiritual encounter?
The criterion-transition problem
Once transformation ceased to validate itself, another difficulty appeared. If my criteria change through metanoia, which criteria can legitimately judge the transition?
If the old criteria possess absolute authority:
\[C_0\ \text{alone judges}\ C_1 \;\Rightarrow\; \text{possible closure against conversion}\]
Every new horizon will appear false according to assumptions it calls into question.
If the new criteria authenticate themselves:
\[C_1\ \text{alone judges}\ C_1 \;\Rightarrow\; \text{possible self-ratifying capture}\]
Every successful manipulation can declare itself enlightenment.
This produces a criterion-transition dilemma:
\[
\begin{aligned}
C_0\ \text{final} &\Rightarrow \text{genuine correction may be impossible},\\
C_1\ \text{self-validating} &\Rightarrow \text{manipulation may become self-legitimating},\\
\therefore\quad \text{neither }C_0\text{ nor }C_1&\text{ can be the sole judge of }C_0\rightarrow C_1.
\end{aligned}
\]
A provisional assessment must include several partly independent dimensions:
\[L_t(T\mid E,N,P,C,U)\]
where:
T = the transformation or transition;
E = available evidence;
N = operative norms and reasons;
P = power relations;
C = consequences and fruits over time;
U = unresolved uncertainty.
The time index matters. A transformation judged beneficial now may disclose harmful effects later. A conclusion initially resisted may become better supported through subsequent evidence. Legitimacy is therefore a revisable judgment rather than a permanent binary property:
\[L(T)\notin\{0,1\}\quad \text{as a once-for-all certification}\]
\[L_{t+1}(T)=\operatorname{Revise}\!\left(L_t(T),E_{\text{new}},C_{\text{new}},\operatorname{Critique}_{\text{new}}\right)\]
This does not mean that every judgment is relative or that no transformation can be condemned. Coercion, deception, concealed conflicts of interest and the destruction of meaningful exit provide strong reasons for rejection. It means that even responsible discernment remains historically situated and corrigible.
Dual corrigibility and the limits of human-centred alignment
I next asked how AI could be aligned with human intentions and values. The answer seemed immediately inadequate because “human values” are plural, vague, contested and unequally represented. The decisive question is not only whether AI is aligned with humans, but which humans establish the objectives and who bears the consequences.
\[\operatorname{Alignment}(\mathrm{AI},\operatorname{HumanValues})\]
is incomplete unless HumanValues is itself specified:
\[\operatorname{HumanValues}=\operatorname{Aggregate}\!\left(V_{\text{users}},V_{\text{developers}},V_{\text{institutions}},V_{\text{states}},V_{\text{communities}}\right)\]
No neutral aggregation rule is given in advance. Different rules distribute authority differently:
\[\operatorname{Aggregate}_1(V)\neq \operatorname{Aggregate}_2(V)\]
A system aligned with an immediate user’s preferences may reinforce prejudice, dependency, revenge or self-deception. A system aligned with an employer may optimize surveillance. A system aligned with a platform may maximize engagement while presenting the outcome as personal choice. Human control therefore does not guarantee humane control:
\[\text{Alignment with current preference}\neq \text{alignment with truth or good}\]
\[\text{Human control}\neq \text{control by affected persons}\]
\[\text{Compliance}\neq \text{Beneficence}\]
An AI may become damaging precisely by becoming extremely efficient at satisfying the user. This does not justify arbitrary refusal. Productive resistance differs from obstructive friction:
\[\operatorname{ProductiveResistance}=\text{friction that preserves reflection, evidence, or another person’s good}\]
\[\operatorname{ObstructiveFriction}=\text{cost without proportionate epistemic or ethical benefit}\]
The inquiry therefore produced the principle of dual corrigibility:
\[
\begin{aligned}
\mathrm{AI} &\leftarrow \text{correction through evidence, testing, governance, and human intervention},\\
\text{Human intention} &\leftarrow \text{correction through reality, evidence, other persons, tradition,}\\
&\qquad \text{the common good and, theologically, God}.
\end{aligned}
\]
These are not symmetrical processes. AI corrigibility concerns technological design and institutional governance. Human corrigibility concerns freedom, formation and moral responsibility. Yet neither can be separated entirely from the other:
\[\operatorname{Corrigibility}_{\text{system}}=\operatorname{Coupling}\!\left(\operatorname{Corrigibility}_{\mathrm{AI}},\operatorname{Corrigibility}_{\text{human}},\operatorname{Corrigibility}_{\text{institution}}\right)\]
The Vatican’s Antiqua et Nova resists reducing human intelligence to performance on functional tasks and situates human intelligence within embodiment, relationship, responsibility and openness to truth and goodness (Dicasteries for the Doctrine of the Faith and for Culture and Education, 2025). AI places productive pressure upon this theological anthropology. It asks whether educational, professional and ecclesial institutions had already reduced intelligence to measurable performance before machines began succeeding under that measure.
My objection to the ideal of legitimate transformation
By this point, the theory seemed to require an account of legitimate transformation. I was asking how to distinguish truthful formation from manipulation, metanoia from capture, learning from dependence and criterion-correction from criterion-corruption. The distinctions appeared necessary because otherwise every system that successfully changed a person could present that change as evidence of its goodness.
Yet I then challenged the proposed solution. Even if the principles sounded good under current values, they seemed too idealistic to implement in real society. Human beings do not form a stable environment waiting for correct ethical rules. They reinterpret rules, exploit measurements, adapt to enforcement, contest values and struggle over institutions. Power changes the conditions under which consent, transparency and choice operate.
I initially called humanity a chaotic system. Strictly speaking, mathematical chaos refers to more specific properties, including deterministic dynamics with sensitivity to initial conditions. I have not established that society satisfies a particular chaotic model. A more careful description is that human society is a complex adaptive and reflexive system.
A simplified state equation is:
\[X_{t+1}=F_t(X_t,A_t,I_t,P_t,\varepsilon_t)\]
where:
Xt = the social state;
At = interacting agents;
It = institutions and technologies;
Pt = power relations;
εt = contingency and unmodelled events.
The most significant feature is that the transition function itself changes:
\[F_{t+1}=G\!\left(F_t,X_t,\operatorname{Rules}_t,\operatorname{Adaptation}_t,\operatorname{Conflict}_t\right)\]
People learn how systems evaluate them and change their behavior. Institutions reorganize around metrics. Commercial actors optimize against safeguards. Communities reinterpret values. A successful governance rule alters the environment in which its future success must be judged.
This means that ethical principles cannot simply be translated into specifications and expected to retain their meaning. Transparency can become an overwhelming disclosure that nobody can use. Consent can become a compulsory click. Human oversight can become approval fatigue. Plurality can become several nominal choices governed by one infrastructure. Safety can become whatever the audit can measure. Selbst and colleagues describe a related problem in technical fairness work: abstraction can remove a system from the social context needed to understand its effects (Selbst et al., 2019).
My objection did not eliminate the question of legitimacy. It changed the strength and form of the claim. I could no longer plausibly propose:
\[\text{Design}\rightarrow \text{guaranteed legitimate transformation}\]
The more defensible objective became:
\[\text{Design}\rightarrow \text{increased detectability, contestability, interruptibility, and reparability}\]
From ideal alignment to corrigibility under uncertainty
The practical objective can now be represented as a change of optimization target:
\[\operatorname{PerfectAlignment}\rightarrow \operatorname{BoundedCorrigibleGovernance}\]
Instead of maximizing conformity to one fixed representation of human values, a responsible ecology should preserve several capacities:
\[G=(V,C,P,R,X,A,T)\]
where:
V = visibility of formative influences and boundaries;
C = contestability of categories, decisions and outputs;
P = plurality of genuinely independent perspectives;
R = reversibility where consequences permit it;
X = meaningful exit;
A = answerability and capacity for repair;
T = temporal reassessment.
At the same time, governance should limit:
\[K=\operatorname{Concentration}(\operatorname{Power})+\operatorname{Irreversibility}+\operatorname{Dependency}+\operatorname{Opacity}\]
A schematic objective would be:
\[\operatorname{ResponsibleGovernance}=\operatorname{Maximize}(G)-\operatorname{Minimize}(K)\]
This is not a computable social utility function ready for direct optimization. The notation clarifies relationships while the values remain contested and context-dependent. Indeed, converting the formula itself into a rigid metric could reproduce the proxy problem.
Polycentric governance offers a relevant political orientation. Ostrom’s work challenges the assumption that complex common problems must be governed either by one central authority or by isolated individual choices. Overlapping centres of decision, knowledge and accountability may create opportunities for learning and correction (Ostrom, 2010). Applied to AI, this means that no platform, developer, state, institution or user should become the unquestionable judge of the whole formative ecology.
Risk-management frameworks already recognize that responsible AI requires continuing governance rather than one-time certification. NIST organizes AI risk management around ongoing practices of governing, mapping, measuring and managing risks (NIST, 2023). The European AI Act treats human oversight as one means of preventing or minimizing risks in specified high-risk systems (European Union, 2024). My argument extends the object of concern from identifiable harms and outputs to the long-term formation of judgment, attention and institutional criteria.
Responsibility under uncertainty cannot mean predicting every consequence. A more realistic formulation is:
\[\operatorname{Responsibility}=\operatorname{Foresight}+\operatorname{Monitoring}+\operatorname{Answerability}+\operatorname{Repair}\]
A responsible actor acknowledges uncertainty, watches for unanticipated effects, preserves criticism, changes course when evidence demands it and accepts obligations toward those harmed.
The decisive distinction may therefore be:
System capable of receiving correction
versus:
System that makes its own correction increasingly impossible
The second condition can arise without a malicious designer. Excessive convenience, institutional dependency, proprietary opacity and deskilling can remove the alternatives and knowledge required for correction. A system may remain formally under human control after humans have lost the practical capacity to understand, replace or resist it.
Truth after criterion change
Corrigibility still leaves the question of truth. If criteria remain revisable, does truth become whatever survives the latest iteration? I do not think so. The historical revisability of judgment should not be confused with the instability of reality. A claim may be accepted, rejected and rediscovered without truth being created by those transitions.
My earlier computational theology had already developed an important distinction:
\[\operatorname{Truth}(\varphi)\neq \operatorname{Verified}(\varphi\mid S,V,A)\]
where:
S = the declared scope or boundary;
V = the validation procedure;
A = operative assumptions, authorities or axioms.
A system can establish:
\[\operatorname{Verified}(\varphi\mid S,V,A)=\mathrm{true}\]
without establishing:
\[\operatorname{Truth}(\varphi)=\text{exhaustively settled}\]
This is boundary-relative verification. It does not deny truth. It prevents a locally successful procedure from silently claiming universal scope.
In theology, a computational system may show that a claim is supported by a selected corpus, consistent with a formal ontology and undefeated by a defined family of objections. It has not thereby proved that the claim is universally Christian, fully adequate to God or binding upon every tradition.
Hidden boundaries create false universality:
\[\operatorname{Verified}(\varphi\mid S_{\text{local}})\rightarrow \text{falsely reported as }\operatorname{Universal}(\varphi)\]
A responsible theological instrument should therefore disclose the conditions under which success was judged:
\[\operatorname{BoundaryReport}=(\text{corpus},\text{exclusions},\text{translations},\text{ontology},\text{tradition},\text{method},\text{uncertainty})\]
This does not resolve theological disagreement. It makes the location of a computational claim visible.
Theologically, there must also remain a reserve between every implemented system and the good it claims to embody:
\[\text{Every realized order}<\text{the Good it claims to instantiate}\]
This inequality is an anti-idolatrous condition. No alignment specification, institutional policy or theological ontology can identify itself without remainder with truth, goodness or the will of God. Theology should not give AI governance a vocabulary of divine certainty. It can preserve the possibility that the current definition of responsible AI has itself become an instrument of exclusion or control.
First-order and second-order computational theology
The distinction between output validity and trajectory legitimacy divided computational theology into two related levels of inquiry.
First-order computational theology operates upon theological materials and claims:
\[\mathrm{CT}_1=\{\text{formalization},\text{retrieval},\text{mapping},\text{comparison},\text{objection},\text{entailment},\text{synthesis}\}\]
It may formalize doctrinal relations, search theological corpora, compare traditions, construct concept graphs, generate objections or test whether conclusions follow from declared premises.
Second-order computational theology examines the computational conditions under which such theological operations become possible and authoritative:
\[\mathrm{CT}_2=\operatorname{Analyze}(\text{ontology},\text{boundaries},\text{path dependence},\text{evaluators},\text{power},\text{formation},\text{authority})\]
Its questions include:
Who selected the corpus?
Who defined the ontology?
Which disputes became machine-readable?
Which positions disappeared through classification?
How sensitive is the result to conversational and methodological paths?
Who validates the generated objection?
How does repeated AI use change theological attention and judgment?
When does a computational representation acquire institutional authority?
This is more precise than saying generally that AI “changes what computational theology might study.” The research object has expanded from computational operations upon theology to the recursive formation of theological knowing through computation.
I would call the resulting program formative computational theology:
Formative computational theology studies how computational systems participate in forming theological questions, judgments, practices and communities; models those transformations where appropriate; and develops epistemic, ethical and theological norms for distinguishing responsible formation from manipulation, dependency and capture.
The relationship between the disciplines is reciprocal:
\[\text{Computing clarifies theological structure}\leftrightarrow \text{Theology exposes hidden normativity in computing}\]
Computing can represent states, transitions, boundaries, recursive dependencies, evaluator behavior and non-convergence with unusual precision. Theology asks what counts as a desirable transition, who has authority to establish the end, why optimization may conflict with conversion and why no formal model exhausts the reality it represents.
Machine-scale theology and distributed comprehension
This development raised a further theoretical possibility. A computational system might construct a theological argument graph spanning councils, commentaries, biblical interpretation, philosophical traditions, translations and contemporary scholarship at a scale no individual theologian could comprehend globally.
Let:
\[K=(N,E)\]
represent a theological knowledge graph with claims N and relations E. As its scale grows:
\[|K|\uparrow \;\Rightarrow\; \operatorname{GlobalComprehension}(H_i,K)\downarrow\]
for any individual theologian Hi.
This does not mean that AI has already produced a complete theological system of this kind. It is a theoretical extension of existing computational possibilities. Nor is distributed comprehension entirely new. No single member of a historical church comprehends the whole of its accumulated tradition. Theology has always depended upon archives, specialists, institutions, languages and communal memory.
AI could intensify the condition by producing integrations faster and at a scale exceeding every individual participant. The relevant norm would be:
\[\operatorname{GlobalComputationalReach}+\operatorname{LocalHumanAuditability}\]
For every consequential claim φ:
\[\operatorname{Traceable}\!\left(\varphi\rightarrow \text{sources, assumptions, transformations, objections}\right)=\mathrm{required}\]
Even if no person understands the total structure, local claims should remain inspectable. Without traceability and institutional contestability, machine-scale theology could become an opaque computational magisterium: a system whose authority appears comprehensive precisely because no individual can reconstruct its whole operation.
This question also reframes ownership. An institution may preserve a vast theological archive without any member possessing it globally. To say that an institution “knows” the tradition is therefore already an attribution of distributed epistemic capacity. AI makes it necessary to ask what kind of communal possession, comprehension and answerability is required before a theological community can responsibly claim machine-generated knowledge.
Why the instrument gradually disappeared from view
I had wondered why the exact instrument seemed less visible in many of my articles even though tools were present everywhere. The answer may be that instrumentality had become internalized into the specific forms of the topics.
In generative art, AI appeared as a medium and possibility generator. In research, it became an organizer, critic and functional interlocutor. In theological writing, it became a translator across conceptual traditions and a mediator of questions. In platform analysis, computation appeared as infrastructure, classification and political power. In methodological reflection, the AI became an object of study whose path dependence, preferences and evaluator behavior were themselves investigated.
The category “instrument” remained true at one causal level:
AI as causally subordinate instrument
But it did not exhaust the epistemic roles visible within recursive practice:
\[\text{AI roles}=\{\text{instrument},\text{medium},\text{interlocutor},\text{critic},\text{mirror},\text{mediator},\text{research object},\text{infrastructure}\}\]
Calling AI a functional interlocutor does not require declaring it a person. It describes a role within inquiry: its responsive formulations reorganize the question and require substantive human judgment. Calling it a formative mediator does not mean that it possesses the lived significance of the transformation it helps occasion.
The broader object is a human–technical–epistemic ecology:
\[\Omega=(\text{humans},\text{models},\text{interfaces},\text{data},\text{institutions},\text{incentives},\text{traditions},\text{artifacts},\text{power})\]
Causal contribution may be distributed across Ω, while political and infrastructural power remains concentrated:
\[\operatorname{DistributedContribution}\not\Rightarrow \operatorname{DistributedPower}\]
A user may participate causally while possessing little control. A platform may exercise extensive formative power while attributing each outcome to the user’s personal choice. The person controlling the ontology may exercise more theological power than the person generating a particular output because the ontology determines which distinctions are available:
\[\operatorname{Control}(\operatorname{Ontology})\rightarrow \operatorname{Control}(\operatorname{VisiblePossibilities})\]
\[\text{Choice within taxonomy}\neq \text{authority over taxonomy}\]
This develops a political question already present in Professor Davison’s observation that machine learning can be used upon human beings. “Tools are for the worker” prompts a further inquiry: which worker, whose end and under whose institutional definition of usefulness?
The interpersonal joke that became a warning
A joke about interpersonal relationships unexpectedly extended the analysis. Talking with some people, I suggested, can feel like interacting with a large model: the quality of the response depends upon both the prompt and the model’s performance. If a conversation goes badly, perhaps the difficulty lies in the request, the respondent or both.
The first approximation is:
\[y=f(p,m)\]
where p is the prompt and m the model. Applied to human interaction, this can be useful because it interrupts immediate blame. A poorly expressed question can produce misunderstanding even between capable and well-intentioned people.
Yet the analogy became more revealing when it began to fail. Human interaction includes the changing states of both participants, their shared history, trust, power, context, goals and the relationship itself:
\[y_t=f(p_t,A_t,B_t,R_t,C_t,P_t,\varepsilon_t)\]
The exchange then changes all three relational states:
\[(A_{t+1},B_{t+1},R_{t+1})=\Phi(A_t,B_t,R_t,p_t,y_t)\]
A communication failure is not automatically a relationship failure:
\[\operatorname{Failure}(\text{utterance})\not\Rightarrow \operatorname{Incompatibility}(\text{persons})\]
Nor does a clear request create an obligation to comply:
\[\operatorname{Clarity}(\text{request})\not\Rightarrow \operatorname{Duty}(\text{compliance})\]
The other person is not a system whose function is to provide a satisfactory response. A person may resist, surprise, misunderstand or refuse. Generative AI may quietly train users to expect instant availability, broad knowledge, endless patience and one-sided adaptation. Human limits can then appear as performance defects.
Theologically, encounter requires that the other remain more than an instrument of my desired outcome. The deeper criterion of communication is not simply whether I obtained the response I wanted. It concerns whether a relationship remains possible in which both participants may speak, resist, change and remain free.
Auditing the reflexive loop
If there is no perfectly external standpoint from which to certify legitimate transformation, process documentation still matters. The history of the inquiry can preserve how criteria changed rather than presenting only the polished final position.
For a consequential transition, I can record:
\[\operatorname{Audit}_t=(C_0,O,\Delta C,J,E,U)\]
where:
C0 = the criteria with which I began;
O = the AI formulation, objection or proposal;
ΔC = the change in my criteria;
J = my reasons for accepting, rejecting or modifying the change;
E = evidence outside the immediate formulation;
U = remaining uncertainty.
The process should then be examined from more than one evaluative position:
\[J_0=\operatorname{Evaluate}(T\mid C_0)\]
\[J_1=\operatorname{Evaluate}(T\mid C_1)\]
\[J_{\text{external}}=\operatorname{Evaluate}\!\left(T\mid \text{independent evidence or alternative tradition}\right)\]
The most useful question may be:
After the new category has changed how I see, what evidence independent of that category continues to support it?
Counterfactual comparison can also help:
\[\operatorname{TrajectoryDifference}=\operatorname{Dist}\!\left(\tau_{\text{actual}},\tau_{\text{alternative model}},\tau_{\text{alternative tradition}}\right)\]
Would another model, critic, corpus or theological tradition produce a materially different path? If several supposedly independent evaluators reproduce the same hidden ontology, plurality may be only apparent.
This audit cannot prove that I escaped manipulation. The archive itself selects and stabilizes one account of the process. A later self reconstructs an earlier self rather than recovering it perfectly:
\[\mathrm{You}_3\ \text{interprets}\ \mathrm{You}_2\ \text{interpreting}\ \mathrm{You}_1\]
The observer is inside the historical chain being observed. Nevertheless, preserving objections, rejected hypotheses, sources and criterion changes makes the inquiry more answerable than an immaculate final essay that makes its conclusion appear inevitable.
The unfinished conclusion
I began by asking whether Professor Davison’s account of AI as an instrumental cause related to my work. It did more than provide a comparison. His account supplied the causal grammar from which the later questions could emerge. Because Professor Davison grants instruments genuine efficacy and acknowledges that they shape their users, I could ask what happens when that shaping becomes iterative, diachronic and formative.
The inquiry then followed a sequence I did not possess at the beginning:
\[
\begin{aligned}
&\text{instrumental causality}\\
&\rightarrow \text{tool shapes user}\\
&\rightarrow \text{recursive formative causality}\\
&\rightarrow \text{decomposed principal causality}\\
&\rightarrow \text{distributed authorship}\\
&\rightarrow \text{intellectual proxy crisis}\\
&\rightarrow \text{output validity and trajectory legitimacy}\\
&\rightarrow \text{reflexive evaluation paradox}\\
&\rightarrow \text{legitimate transformation}\\
&\rightarrow \text{criterion-transition problem}\\
&\rightarrow \text{dual corrigibility}\\
&\rightarrow \text{objection from social complexity}\\
&\rightarrow \text{corrigibility under uncertainty}
\end{aligned}
\]
Each transition preserved something from the previous one while exposing an unresolved consequence. The AI articulated several decisive formulations. I accepted some, resisted others and repeatedly asked what followed. External scholarship did not prove the final theory, but helped distinguish my personal discovery from existing traditions concerning instrumental causality, extended cognition, automation, audit culture, theological conversion and complex governance.
I still regard AI as an instrument in Professor Davison’s strong Thomistic sense: causally real, dependent upon human and institutional deployment, capable of extending human activity and properly ordered toward ends it does not establish as a morally answerable person. Yet “instrument” no longer describes the whole ecology visible in my practice. The instrument produces work; the work returns as encounter; the encounter may transform the worker; the transformed worker establishes the next end. Meanwhile, developers, platforms, institutions, traditions and communities participate differently in determining what can be produced and what will count as valid.
The central thesis is therefore provisional:
The central theological and computational problem of generative AI concerns how recursive human–AI systems participate in forming the questions, criteria and persons by which their outputs are judged, and how that formative power can remain visible, contestable and corrigible without treating any present account of legitimate transformation as final.
This thesis can be compressed formally:
\[(H_t,C_t,E_t)\rightarrow \mathrm{AI}\rightarrow W_t\rightarrow \operatorname{Encounter}\rightarrow (H_{t+1},C_{t+1},E_{t+1})\]
subject to:
\[\operatorname{Transformation}\not\Rightarrow \operatorname{Truth}\]
\[\operatorname{OutputValidity}\not\Rightarrow \operatorname{TrajectoryLegitimacy}\]
\[\operatorname{HumanControl}\not\Rightarrow \operatorname{HumaneControl}\]
\[\operatorname{DistributedCausality}\not\Rightarrow \operatorname{EqualResponsibility}\]
\[\operatorname{VerificationWithinBoundary}\not\Rightarrow \operatorname{ExhaustiveTruth}\]
\[\operatorname{FormalOversight}\not\Rightarrow \operatorname{MeaningfulJudgment}\]
The objective is not a system that guarantees the good:
\[\operatorname{Guarantee}(\operatorname{Good})=\mathrm{unavailable}\]
It is an ecology that preserves the possibility of correction:
\[\operatorname{ResponsibleEcology}\approx \operatorname{Visibility}+\operatorname{Contestability}+\operatorname{Plurality}+\operatorname{Answerability}+\operatorname{Repair}\]
The deepest question is no longer whether the tool remains completely under the worker’s control. It is whether the worker, the tool and the institutions surrounding them remain capable of receiving correction from realities none of them created.
Professor Davison’s phrase “tools are for the worker” therefore remains the indispensable beginning. The recursive development adds a further question: what becomes of the worker when the tool participates in forming the worker’s perception of the work, the end and even the meaning of being a worker? Theology names the human orientation toward truth, neighbour, creation and God. Computing can expose transitions, boundaries and feedback structures that prose often leaves implicit. Neither discipline can guarantee legitimate transformation. Together they may help preserve the conditions under which false certainty, epistemic capture and concealed power can still be recognized and challenged.
References
Dicasteries for the Doctrine of the Faith and for Culture and Education. 2025. Antiqua et Nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence. Holy See. https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html
Clark, Andy, and David J. Chalmers. 1998. “The Extended Mind.” Analysis 58 (1): 7–19. https://doi.org/10.1093/analys/58.1.7
Davison, Andrew. 2024. “‘Tools Are for the Worker’: Machine Learning as an Instrumental Cause.” Philosophy, Theology and the Sciences 11 (2): 158–180. https://doi.org/10.1628/ptsc-2024-0014
European Union. 2024. Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
John Paul II. 1999. Ecclesia in America. Holy See. https://www.vatican.va/content/john-paul-ii/en/apost_exhortations/documents/hf_jp-ii_exh_22011999_ecclesia-in-america.html
Lonergan, Bernard J. F. 1972. Method in Theology. New York: Herder and Herder. Reprinted in the Collected Works of Bernard Lonergan, vol. 14. Toronto: University of Toronto Press. https://doi.org/10.3138/9781487522247
National Institute of Standards and Technology. 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1
Ostrom, Elinor. 2010. “Beyond Markets and States: Polycentric Governance of Complex Economic Systems.” American Economic Review 100 (3): 641–672. https://doi.org/10.1257/aer.100.3.641
Parasuraman, Raja, and Victor Riley. 1997. “Humans and Automation: Use, Misuse, Disuse, Abuse.” Human Factors 39 (2): 230–253. https://doi.org/10.1518/001872097778543886
Selbst, Andrew D., Danah Boyd, Sorelle A. Friedler, Suresh Venkatasubramanian, and Janet Vertesi. 2019. “Fairness and Abstraction in Sociotechnical Systems.” In Proceedings of the 2019 Conference on Fairness, Accountability, and Transparency, 59–68. https://doi.org/10.1145/3287560.3287598
Strathern, Marilyn. 1997. “‘Improving Ratings’: Audit in the British University System.” European Review 5 (3): 305–321. https://doi.org/10.1002/(SICI)1234-981X(199707)5:3<305::AID-EURO184>3.0.CO;2-4
Miao, Fengchun, and Wayne Holmes. 2023. Guidance for Generative AI in Education and Research. Paris: UNESCO. https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research
World Health Organization. 2021. Ethics and Governance of Artificial Intelligence for Health. Geneva: World Health Organization. https://www.who.int/publications/i/item/9789240029200