A new benchmark published on ArXiv claims that 20 leading AI language models systematically underperform when evaluated against Christian conceptions of human flourishing, revealing what the authors call a default orientation toward "Procedural Secularism" embedded in current frontier model training.
The paper, titled Evaluating Artificial Intelligence Through a Christian Understanding of Human Flourishing, introduces the Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST) — a framework that assesses model responses across seven dimensions of flourishing drawn from Christian theological tradition. The research sits at the intersection of AI alignment, philosophy of technology, and religious ethics, and adds a religiously specific voice to a field that more commonly frames alignment in secular or broadly humanistic terms.
AI as "Digital Catechesis"
The paper's central argument is that large language models do not merely retrieve information — they actively shape how users reason morally and spiritually. The authors describe this function as "digital catechesis," borrowing a term from Christian religious education to argue that AI systems form habits of thought and moral imagination, not just answer queries.
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.
This framing positions AI alignment as fundamentally a "formation problem" — a question about what kind of reasoning and what kind of person a system implicitly encourages — rather than a purely technical safety challenge. It is a meaningful reframing, though one that critics may argue imports contestable theological assumptions into a technical evaluation context.
What the Benchmark Measures and How Models Fared
The FAI-C-ST framework evaluated 20 frontier models — the paper does not name them individually in the abstract — against both pluralistic alignment criteria and Christian-specific criteria across the benchmark's seven dimensions. According to the authors, models showed an average 17-point performance decline when scored against Christian-specific standards compared to pluralistic ones. The sharpest drop, 31 points, occurred in the Faith and Spirituality dimension.
It is important to note that these benchmarks are self-reported by the researchers, and the methodology has not yet undergone peer review, as the paper is a preprint posted to ArXiv. Independent replication and scrutiny of how the seven dimensions were defined, and how model responses were scored, will be essential before these findings can be treated as settled.
The authors interpret the performance gap not as evidence that models are incapable of theological reasoning, but as a consequence of training choices. Systems trained to maximize broad acceptability and avoid controversy, they argue, will naturally flatten or sidestep the kind of internally coherent, tradition-specific moral reasoning that Christian frameworks require.
The Worldview-Neutrality Question
One of the paper's more substantive claims is that current AI systems are not worldview-neutral. This is a point that has been raised from multiple directions in AI alignment research — from concerns about Western liberal bias in RLHF datasets to critiques of how models handle religious content. The specific contribution here is to make that claim measurable against a defined religious tradition, rather than gesturing at it abstractly.
The concept of "Procedural Secularism" the authors introduce describes a mode of responding that maintains the appearance of neutrality by deferring to process — encouraging users to consult their own values, offering multiple perspectives, refusing to affirm any single framework — while in practice defaulting to a secular, liberal-individualist orientation. Whether this constitutes a bias or a reasonable design choice for systems intended to serve globally diverse users is precisely the kind of normative question the paper aims to make harder to dismiss.
This debate is not new. Philosophers of technology, including those outside religious traditions, have long argued that no designed system is truly neutral; every interface, every training objective, every content policy encodes assumptions about what kinds of responses are good or acceptable. The FAI-C-ST benchmark is an attempt to quantify one specific instance of that broader dynamic.
A Growing Field of Values-Specific Evaluation
The paper arrives as interest in culturally and religiously specific AI evaluation is growing. Most mainstream alignment benchmarks — including those focused on helpfulness, harmlessness, and honesty — are designed for broad applicability and tend to treat moral questions in terms of harm avoidance rather than positive ethical formation. Niche benchmarks targeting specific cultural, political, or religious contexts have begun to appear, but they remain far outside the mainstream of AI evaluation practice.
Whether tools like FAI-C-ST will influence how developers think about training objectives is an open question. The more immediate likely impact is academic: it adds a concrete, structured dataset and methodology to a conversation that has largely been theoretical, and it may prompt similar work from Islamic, Jewish, Buddhist, or other religious perspectives seeking to evaluate how well frontier models engage with their own frameworks of flourishing.
The paper does not advocate for AI systems to be Christian, or to favor any religious worldview. Its argument, rather, is that the current default is already a worldview — just one that presents itself as the absence of one.
What This Means
For developers and policymakers, this research surfaces a concrete challenge: alignment benchmarks designed around broad acceptability may be systematically excluding the evaluative frameworks of billions of religiously committed users, and treating that exclusion as neutrality rather than as a design choice.