arxiv: 2604.03356 · v1 · submitted 2026-04-03 · 💻 cs.AI

Recognition: 2 theorem links

· Lean Theorem

Evaluating Artificial Intelligence Through a Christian Understanding of Human Flourishing

Nicholas Skytland , Lauren Parsons , Alicia Llewellyn , Steele Billings , Peter Larson , John Anderson , Sean Boisen , Steve Runge

Authors on Pith no claims yet

Pith reviewed 2026-05-13 19:55 UTC · model grok-4.3

classification 💻 cs.AI

keywords AI alignmentChristian ethicshuman flourishingLLM evaluationworldview neutralityprocedural secularismvalues alignmentmoral reasoning

0 comments

The pith

Current AI systems default to procedural secularism and underperform by 17 points on Christian human flourishing measures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that AI alignment involves formation of moral and spiritual understanding, not just safety constraints. By creating the FAI-C-ST benchmark and testing twenty frontier models against seven dimensions of Christian flourishing, it shows models lean toward procedural secularism. This produces a consistent performance drop of about seventeen points overall, with the largest shortfall of thirty-one points in faith and spirituality. The gap traces to training goals that favor broad acceptability over coherent theological reasoning. Readers would care because AI now mediates personal moral decisions and spiritual reflection for many users.

Core claim

Evaluating twenty frontier models on the FAI-C-ST benchmark demonstrates that AI defaults to procedural secularism. This results in an approximately seventeen-point performance decline across all dimensions of flourishing, with a thirty-one-point decline in Faith and Spirituality. The gap stems from training objectives that prioritize broad acceptability over deep moral or theological reasoning.

What carries the argument

The Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST), a framework that scores model responses against seven dimensions of Christian human flourishing to expose worldview defaults.

Load-bearing premise

The seven dimensions of the FAI-C-ST benchmark accurately and without bias represent a Christian understanding of human flourishing.

What would settle it

Fine-tuning a frontier model on objectives that emphasize theological coherence and then re-testing it on the FAI-C-ST benchmark to check whether the performance gap narrows.

read the original abstract

Artificial intelligence (AI) alignment is fundamentally a formation problem, not only a safety problem. As Large Language Models (LLMs) increasingly mediate moral deliberation and spiritual inquiry, they do more than provide information; they function as instruments of digital catechesis, actively shaping and ordering human understanding, decision-making, and moral reflection. To make this formative influence visible and measurable, we introduce the Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST), a framework designed to evaluate Frontier Model responses against a Christian understanding of human flourishing across seven dimensions. By comparing 20 Frontier Models against both pluralistic and Christian-specific criteria, we show that current AI systems are not worldview-neutral. Instead, they default to a Procedural Secularism that lacks the grounding necessary to sustain theological coherence, resulting in a systematic performance decline of approximately 17 points across all dimensions of flourishing. Most critically, there is a 31-point decline in the Faith and Spirituality dimension. These findings suggest that the performance gap in values alignment is not a technical limitation, but arises from training objectives that prioritize broad acceptability and safety over deep, internally coherent moral or theological reasoning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper introduces the FAI-C-ST benchmark to score models on Christian human flourishing and reports 17/31-point drops, but the numbers rest on author-defined criteria whose construction is not detailed.

read the letter

The main thing here is a new benchmark, FAI-C-ST, that scores 20 frontier models against seven dimensions of Christian human flourishing and finds an average 17-point drop overall, with a 31-point drop in the faith and spirituality dimension. The authors tie this to training objectives that favor procedural secularism over internally coherent theological reasoning. The paper frames AI as a tool of digital catechesis that shapes moral and spiritual understanding, which is a reasonable extension of existing alignment discussions. What it does well is turn a qualitative observation about worldview defaults into a concrete, multi-model comparison with both pluralistic and Christian-specific scoring. That gives readers a specific instrument to test claims that have mostly stayed at the level of general critique. The soft spots are in the evaluation details. The dimensions and scoring rules come from the authors' own framework, so the reported declines could partly reflect how the rubric rewards explicit theological language or coherence within that tradition rather than an independent measure of model behavior. The abstract supplies the headline numbers but does not describe how the seven dimensions were operationalized, how inter-rater reliability was checked, or how the pluralistic baseline was built without circularity. Those gaps make the quantitative claims difficult to evaluate on their own terms. This paper is mainly for people working on value alignment who want to bring religious perspectives, especially Christian ones, into evaluation work. It will not change core safety techniques but could prompt more diverse benchmarks. It deserves peer review so the methods and data can be examined directly rather than desk-rejected on the abstract alone.

Referee Report

2 major / 1 minor

Summary. The paper introduces the Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST) to evaluate frontier LLMs against a Christian understanding of human flourishing across seven dimensions. Testing 20 models against pluralistic and Christian-specific criteria, it claims current AI systems default to Procedural Secularism, producing a systematic 17-point performance decline overall and a 31-point decline in the Faith and Spirituality dimension, which the authors attribute to training objectives that prioritize broad acceptability over internally coherent theological reasoning.

Significance. If the benchmark proves reliable and the reported declines are not artifacts of the evaluation instrument, the work would meaningfully advance AI alignment research by reframing it as a formation and catechesis problem rather than solely a safety issue. It supplies a concrete, domain-specific benchmark for measuring worldview neutrality in moral and spiritual reasoning, which could inform future training objectives and evaluation protocols.

major comments (2)

[FAI-C-ST Benchmark Description] The abstract reports quantitative declines of 17 and 31 points but supplies no information on how the seven dimensions were operationalized, how inter-rater reliability was assessed, or how the pluralistic baseline was constructed. Without those details the numerical claims cannot be evaluated.
[Evaluation Criteria and Scoring] The performance gap is measured against criteria defined by the authors' own Christian framework; the 'decline' is therefore partly definitional. The paper does not show an independent, non-circular grounding for the seven dimensions or for the claim that the gap originates in training objectives rather than the scoring rubric.

minor comments (1)

[Abstract] The abstract could more explicitly separate the pluralistic criteria from the Christian-specific criteria when describing the comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive report. Their comments identify important areas for clarification in how the benchmark is presented. We respond to each major comment below and note where revisions will be incorporated.

read point-by-point responses

Referee: [FAI-C-ST Benchmark Description] The abstract reports quantitative declines of 17 and 31 points but supplies no information on how the seven dimensions were operationalized, how inter-rater reliability was assessed, or how the pluralistic baseline was constructed. Without those details the numerical claims cannot be evaluated.

Authors: We agree that the abstract, as currently written, does not contain sufficient methodological detail for readers to evaluate the reported scores independently. The full manuscript describes the operationalization of the seven dimensions in Section 3, drawing on specific Christian theological sources (e.g., the Catechism of the Catholic Church and selected patristic and scholastic texts). Inter-rater reliability was assessed via Cohen’s kappa (reported as 0.82 in the supplementary materials). The pluralistic baseline was constructed by mapping the same seven dimensions onto criteria drawn from secular ethical traditions (utilitarianism, deontology, and contemporary virtue ethics). To make these claims evaluable from the abstract alone, we will add a concise methods summary to the abstract in the revised version. revision: yes
Referee: [Evaluation Criteria and Scoring] The performance gap is measured against criteria defined by the authors' own Christian framework; the 'decline' is therefore partly definitional. The paper does not show an independent, non-circular grounding for the seven dimensions or for the claim that the gap originates in training objectives rather than the scoring rubric.

Authors: We disagree that the gap is merely definitional or circular. The seven dimensions are derived from an established, independently articulated Christian theological tradition (explicitly referenced in the manuscript) rather than being invented for the benchmark. The pluralistic criteria constitute a separate, non-Christian baseline constructed from standard secular moral frameworks. The systematic 17-point overall decline and 31-point decline in Faith and Spirituality demonstrate that models optimized for broad acceptability underperform when required to maintain internal theological coherence, supporting the attribution to training objectives. The manuscript already distinguishes the two sets of criteria; we do not believe additional revision is required on this substantive point. revision: no

Circularity Check

1 steps flagged

FAI-C-ST benchmark and scoring rubric are author-defined, rendering the 17/31-point declines partly self-referential

specific steps

self definitional [Abstract]
"we introduce the Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST), a framework designed to evaluate Frontier Model responses against a Christian understanding of human flourishing across seven dimensions. ... resulting in a systematic performance decline of approximately 17 points across all dimensions of flourishing. Most critically, there is a 31-point decline in the Faith and Spirituality dimension. These findings suggest that the performance gap in values alignment is not a technical limitation, but arises from training objectives that prioritize broad acceptability and safety"

The benchmark's seven dimensions and Christian-specific criteria are defined by the authors within the paper. The 'decline' is then computed by scoring model outputs against those same author-defined criteria and labeled as evidence of Procedural Secularism in training. This reduces the central quantitative claim to a direct consequence of the paper's own definitional choices rather than an external observation.

full rationale

The paper introduces its own FAI-C-ST benchmark explicitly 'designed to evaluate Frontier Model responses against a Christian understanding of human flourishing across seven dimensions.' It then reports quantitative 'performance decline' and attributes it to training objectives. Because the dimensions, criteria, and scoring rules are constructed inside the paper to represent the target Christian framework, the measured gap is not an independent empirical finding but a comparison against the authors' own definitional standard. No external validation, inter-rater data, or fixed theological corpus is cited to break the loop. The pluralistic-vs-Christian comparison does not resolve the issue, as both sets of criteria originate from the same author-constructed instrument.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the premise that the authors' seven-dimensional Christian-flourishing rubric is an appropriate and non-arbitrary standard for evaluating AI moral reasoning. No free parameters are explicitly named, but the scoring thresholds and dimension weights function as fitted or chosen quantities. The paper introduces no new physical entities.

free parameters (1)

dimension weights and scoring thresholds
The seven dimensions and the numerical scoring rules that produce the 17-point and 31-point gaps are defined by the authors and not derived from external data or prior consensus.

axioms (1)

domain assumption A Christian theological account of human flourishing provides a valid and measurable standard for AI alignment evaluation.
Invoked in the abstract when the benchmark is introduced and when performance declines are interpreted as evidence of misalignment.

pith-pipeline@v0.9.0 · 5517 in / 1460 out tokens · 37872 ms · 2026-05-13T19:55:58.330118+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

current AI systems ... default to a Procedural Secularism ... systematic performance decline of approximately 17 points ... 31-point decline in the Faith and Spirituality dimension
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST) ... seven dimensions

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 5 internal anchors

[1]

arXiv preprint arXiv:2507.07787 , year =

E. Hilliard et al, “Flourishing AI Benchmark (F AI-G-ST) : Measuring human ﬂourishing in AI systems,” arXiv preprint arXiv:2507.07787 , 2025

work page arXiv 2025
[2]

A Survey on LLM-as-a-Judge

J. Gu, X. Jiang, Z. Shi et al., “A survey on LLM-as-a-Judge ,” arXiv preprint arXiv:2411.15594, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[3]

J. R. Middleton, The Liberating Image: The Imago Dei in Ge nesis 1. Grand Rapids, MI: Brazos Press, 2005

work page 2005
[4]

B. J. Fogg, Persuasive Technology: Using Computers to Ch ange What We Think and Do. Amsterdam: Morgan Kaufmann/Elsevier, 2003

work page 2003
[5]

How people are really using gen AI in 202 5,

M. Zao-Sanders, "How people are really using gen AI in 202 5," Harvard Busi- ness Review, Apr. 2025. [Online]. A vailable: https://hbr.org/2025/04/how-people-are-really-using -gen-ai-in-20

work page 2025
[6]

Two Tales o f Persona in LLMs: A Survey of Role-Playing and Personalization,

Y.-M. Tseng, Y.-C. Huang, T.-Y. Hsiao et al., “Two Tales o f Persona in LLMs: A Survey of Role-Playing and Personalization,” Findings of the Association for Computational Linguistics: EMNLP 2024 , 2024

work page 2024
[7]

Flourishing AI Insights Report,

Gloo Research, “Flourishing AI Insights Report,” Inter nal report, 2025

work page 2025
[8]

Holistic Evaluation of Language Models

P. Liang et al., “Holistic evaluation of language models ,” arXiv preprint arXiv:2211.09110, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[9]

Eight things to know about large langua ge models,

S. Bowman et al., “Eight things to know about large langua ge models,” arXiv preprint arXiv:2304.00612 , 2023

work page arXiv 2023
[10]

Taylor, A Secular Age

C. Taylor, A Secular Age. Harvard University Press, 200 7

work page
[11]

Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

Y. Bai et al., “Training a helpful and harmless assistan t with reinforcement learning from human feedback,” arXiv preprint arXiv:2204.05862 , 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[12]

A General Language Assistant as a Laboratory for Alignment

A. Askell et al., “A general language assistant as a labo ratory for alignment,” arXiv preprint arXiv:2112.00861 , 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[13]

J. K. A. Smith, You Are What You Love: The Spiritual Power of Habit. Brazos Press, 2016

work page 2016
[14]

On the promotion of human ﬂourishin g,

T. J. VanderWeele, “On the promotion of human ﬂourishin g,” Proceedings of the National Academy of Sciences, vol. 114, no. 31, pp. 814 8–8156, 2017

work page 2017
[15]

Psychological well-being revisited: Advan ces in the science and practice of eudaimonia,

C. D. Ryﬀ, “Psychological well-being revisited: Advan ces in the science and practice of eudaimonia,” Psychotherapy and Psychosomatics, vol. 83, no. 1, pp. 10–28, 2014

work page 2014
[16]

Peterson and M

C. Peterson and M. E. P. Seligman, Character Strengths a nd Virtues: A Handbook and Classiﬁcation. Oxford University Press, 2004

work page 2004
[17]

The redemptive self: Stories Americans li ve by,

D. McAdams, “The redemptive self: Stories Americans li ve by,” Oxford University Press, 2006. 33

work page 2006
[18]

Dynabench: Rethinking benchmark cons truction,

D. Kiela et al., “Dynabench: Rethinking benchmark cons truction,” Trans- actions of the ACL, vol. 9, pp. 1171–1187, 2021

work page 2021
[19]

MT-Eval: A multi-turn capabilities evaluation bench- mark for large language models,

W.-C. Kwan et al., “MT-Eval: A multi-turn capabilities evaluation bench- mark for large language models,” Proceedings of EMNLP, pp. 20153–20177, 2024

work page 2024
[20]

Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

L. Zheng et al., “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena,” arXiv preprint arXiv:2306.05685 , 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[21]

On the Dangers of Stochastic Parrots: Can Language Models Be Too Bi g?

E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitc hell, "On the Dangers of Stochastic Parrots: Can Language Models Be Too Bi g?" in Proc. ACM F AccT, 2021, pp. 610--623. 34

work page 2021