Design
We adopted a sequential qualitative-to-quantitative mixed-methods design. The reason for the choice is that while purely qualitative approaches provide interpretive context, they are less suited for evaluating structural frequency patterns or proportional distribution across extensive textual datasets.
Researcher Backgrounds and Task Allocation
Task allocation between the human researchers was strictly structured around their respective domain expertise and methodological backgrounds:
David Leech: Former Senior Lecturer in the Department of Religion and Theology at the University of Bristol; PhD, Faculty of Divinity (Philosophy of Religion), with expertise in qualitative research.
Alexandra Cheung: PhD student in Psychology at the University of Bristol; PGT Facilitator for MSc Applied Neuropsychology and Teaching Assistant for MSc Psychology (Conversion) Research Methods; BSc (Hons) Psychology, University of Sheffield, with experience in mixed-methods research.
To maintain full transparency, the researchers' surnames (Leech and Cheung) are referenced throughout the following methodology section to indicate specific responsibility for each task.
Sampling of Documents
The final analytical corpus comprises 15 primary texts selected across five major Abrahamic traditions (4 Catholic, 4 Protestant, 3 Islamic, 2 Jewish, and 2 Eastern Orthodox). Four documents originating from Buddhist and Hindu traditions were excluded, as it proved methodologically challenging to construct a unified codebook that could meaningfully bridge Abrahamic and non-Abrahamic theological structures alongside the lack of comparable centralized, top-tier institutional statements. To map the institutional standing of each source, the dataset includes descriptive metadata capturing tradition-relative authority tiers (3 levels) alongside discursive genre tags.
Detailed source mappings are catalogued in Appendix A.
Variables
Variables were initially derived deductively by Leech by adapting secular AI safety frameworks (Kulveit et al.) to religious discourse.
Domains (5 Domains with 13 Sub-domains)
Three domain variables were derived from the main societal systems outlined in the paper (D1 Economy, D2 Culture and D3 Nation-State/Politics) as well as two extra domain variables relevant to theological and philosophical documents, namely D4 Humanhood and D5 AI as Conceptual Subject, relevant to theological, philosophical and normative institutional documents. These were subsequently expanded into granular sub-codes (e.g., Within D1, D1a was about macro-allocation of capital while D1b was about the metrics of human labour) by Cheung deductively following a pilot analysis of approximately 100 data extracts.
Moves (6 Moves with 11 Sub-moves)
Similarly, rhetorical move variables (Moves) were deductively structured by Leech to capture the rhetorical trajectory of institutional claims, adapting Swalesian Genre and Move Analysis to normative and theological jurisprudence and progressing from diagnostic to foundational principle to prohibition/norm to call to action. Just like the Domains, these were also subsequently expanded into granular sub-codes by Cheung (e.g. Within M1 Diagnostics, M1a was neutral, M1b was positive and M1c was negative) following the pilot.
Risk Vector variables (4) and Mechanisms (Binary)
Furthermore, Risk Vector variables operationalized Kulveit’s taxonomy of disempowerment pathways, including the core concepts of Gradual Disempowerment (GD) as well as comparative scenarios of Acute Agentic Misalignment (AM), Misuse of Governance (MG) and Power Capture (PC). In addition, a binary variable of Mechanism Presence was adopted to capture the casual mechanism (how) as opposed to end-state descriptions (what).
Policy Demands (5 Categories and 22 Sub-categories) and Stakeholders (5)
Post-hoc coding was conducted by Cheung to code policies, categorizing interventions across five domains of public and regulatory law and operationalizing regulatory instruments drawn from AI legislation and international AI governance frameworks.: Legal Liability (LL), Operational & Engineering Standards (OE), Market & Labor Regulation (ML), Cultural & Digital Sovereignty (CD), and State Defense & Security Limits (SD). They were then expanded into granular sub-codes by Cheung.
Target stakeholders were coded by Cheung to examine if the demand for actions and red lines had a specific target (i.e., tech companies, legislators, governments, regulatory organizations) instead of just the religious community or general humanity.
For the full breakdown of variables, please refer to Appendix B.
Hybrid LLM and Human Pipeline
Overall, the codebook was developed by adapting the deliberative human-LLM collaboration framework proposed by Dai et al. (2023) within a hybrid deductive-inductive process, following a three-stream, hybrid human-in-the-loop pipeline that integrated theoretical derivation with multi-stage automated coding and human auditing.
The workflow began with a methodological setup phase, where initial deductive frameworks (Leech) were expanded through inductive pilot testing on approximately 100 extracts (Cheung using Gemini).
This codebook underwent multi-turn deliberative refinement with Claude to establish explicit boundary definitions, culminating in a gold-standard reference dataset of 50 human-authored exemplars to ground the model before full execution.During primary data processing ($N = 1,790$), automated LLM coding operated under strict human oversight with tiered feedback mechanisms. Extraction rules were continuously monitored by Cheung, with misaligned extracts returned to the model for re-segmentation.
For main coding outputs, human verification pass results branched according to disagreement severity: mild classifications triggered automated re-coding prompts, whereas strong, systematic misalignments (such as Gradual Disempowerment and M3 Redlines) prompted complete manual recoding by Cheung and Leech.
Finally, to handle post-hoc policy and stakeholder code generation without dedicated policy domain expertise, the pipeline deployed a sequential multi-agent audit framework, passing initial automated outputs through two independent, chained Gemini red-team instances to iteratively review and refine final categories.
For a full breakdown of the details of data extraction, codebook creation and refinement and full coding process, please refer to Appendix C.


Data Extraction
The corpus was segmented into data extracts (N = 1790) through a four-stage workflow: 1) initial heuristic rule drafting (Cheung); 2) LLM-assisted linguistic refinement of extraction rules; 3) expert review and ratification (Leech); and 4) automated paragraph-level and intra-paragraph segmentation via LLM (Gemini). Segmentation was strictly governed by explicit rhetorical shift rules, splitting paragraphs whenever arguments transitioned or signpost markers ("Yet," "Therefore") occurred. Each extract was constrained to a 15-45 word unit containing a clear subject, action, and immediate justification. Paragraphs continuously reinforcing a single claim were retained as intact single extracts.
Codebook Development
Overall, the codebook was developed by adapting the deliberative human-LLM collaboration framework proposed by Dai et al. (2023) within a hybrid deductive-inductive process.
An initial codebook draft created by Leech was expanded with an inductive, bottom-up subcode generation using Gemini with human supervision (Cheung) to draft a modified codebook following pilot coding.
It was then followed by a multi-turn deliberative refinement where the human researcher (Leech) and Claude engaged in multi-turn deliberative refinement, such as establishing explicit boundary rules (e.g., D3 vs. D4, M2 vs. M3), changing subcodes and refining definitions until consensus was achieved. Leech then reviewed the corpus and generated 10 initial coding exemplars per tradition to construct a gold-standard reference dataset to ground the model.
To map the religious arguments directly onto policy mechanisms, a final post-hoc codebook refinement was conducted by Cheung. An LLM (Gemini) was utilized to analyze the extracted data and establish standardized policy codes (as well as the granular subcodes) and stakeholder categories in a fully inductive workflow, aligning closely with the original pipeline of Dai et al. (2023), but it did not undergo human-LLM deliberative refinement, as neither researcher had specialized domain expertise in AI policy.
Last but not least, the manual recoding of GD (See below) by Leech led to further refinement of the codebook.
For a summary of the working notes of the deliberative process, please refer to Appendix C.
Hybrid LLM Coding and Human Verification Pipeline
Data annotation was conducted through a hybrid human-LLM pipeline combining initial automated LLM coding with extensive, multi-round manual verification and audit passes by human researchers.
Automated coding was subsequently executed using an LLM coder prompted with both the deliberation JSON of Leech and Claude to establish context. as well as the frozen codebook. Extracts were processed in batches of 10, with fresh model instances initialized every 50 extracts to prevent context-window degradation or attention drift.
Then, multiple rounds of complete manual verification passes were conducted by Cheung and Leech across every data extract to eliminate hallucinations, formatting errors, or classification misalignments. This iterative review included a complete manual recoding of Gradual Disempowerment (GD) by Leech following a systematic misalignment between the model and initial human assumptions, as well as a targeted manual recoding of M3 Redlines by Cheung to resolve coding disagreements. Specifically, during the manual recoding of GD, extracts were graded on a four-tier interpretive levels by Leech which led to further refinement of the codebook.
For policy codes, to further increase validity and mitigate errors stemming from a lack of specialized domain expertise, Cheung deployed two independent instances of Gemini as automated red-team auditors in a sequential pipeline (i.e., Auditor 1 reviewed and modified the output of the LLM Coder, after which Auditor 2 reviewed and refined the output of Auditor 1).