Results
Catholicism
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.
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) resulting in 1790 extracts. 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.
Catholicism
1. Gradual Disempowerment (GD) is Negligible
Explicit GD (GD/y = 6): Just 0.48% of Catholic extracts explicitly articulate Gradual Disempowerment.
Explicit + Boundary (GD/y + GD/y/maybe = 10): Rises to only 0.79%.
Lenient Threshold (Adding GD/maybe = 50): Even including all ambiguous or implicit references (60 total), GD still capped out at 4.75% of the Catholic corpus.
Explicit Rejection (GD/n = 73): Over 7x more extracts explicitly evaluate and miss or turn away from GD (73) than those that explicitly articulate it (6 or 10).
2. Traditional Risk Vectors Dominate
Power Concentration (PC = 105): At 8.32%, Power Concentration is by far the most dominant risk vector in Catholic institutional texts—occurring 17.5x more frequently than explicit GD.
Misuse & Governance (MG = 38): Accounted for 3.01% of extracts (over 6x higher than explicit GD).
Atomic Misalignment (AM = 9): Accounts for 0.71%, showing even basic technical misalignment gets roughly similar explicit airtime as explicit GD (6).
Null / Non-Risk Tagged (null = 977): The vast majority (77.42%) of extracts focus on general ethical framing, human dignity, labor, or pastoral guidance rather than explicit AI threat vector classification.
The Core Finding for the Catholic Section
Overall, Catholic religious leaders demonstrate massive concern over power concentration
monopolization, corporate dominance, and human agency exploitation by central authorities (Power Concentration = 105), but exhibits a near-complete blind spot regarding autonomous loss of control to the system itself (Explicit GD = 6).
Researcher Backgrounds
Task allocation between the human researchers was strictly structured around their respective domain expertise and methodological backgrounds:
David Leech (Researcher 1): Former Senior Lecturer in the Department of Religion and Theology at the University of Bristol
Alexandra Cheung (Researcher 2): PhD student in Psychology at the University of Bristol
To maintain full transparency, the researchers are referenced throughout the following methodology section to indicate specific responsibility for each task.
Hybrid LLM and Human Pipeline
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 (Researcher 1) were expanded through inductive pilot testing on approximately 100 extracts (Researcher 2 using Gemini).
This codebook underwent multi-turn deliberative refinement between Researcher 1 and Claude to establish explicit boundary definitions, culminating in a gold-standard reference dataset of 50 human-authored exemplars (10 per tradition) to ground the model before full execution.
During primary data extraction (N = 1790), automated LLM extraction was conducted under human oversight (Researcher 2) with misaligned extracts returned to the model for re-segmentation.
For the automated coding, human verification branched according to disagreement severity: mild misalignments triggered automated re-coding prompts, whereas strong misalignments (such as Gradual Disempowerment and M3 Redlines) prompted manual recoding by Researcher 1 and Researcher 2.
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, pending expert consultation.
For a full breakdown of the details of data extraction, codebook creation and refinement and full coding process, please refer to Appendix C.

