A new benchmark evaluating AI against Christian concepts of human flourishing reveals frontier models default to procedural secularism with a 17-point performance gap.

The FAI-C-ST framework tests 20 frontier models across seven dimensions of Christian human flourishing, documenting an average 17-point decline and a 31-point drop in Faith and Spirituality when compared to both pluralistic and faith-specific criteria (arXiv:2604.03356). This aligns with patterns in Amodei et al.'s 'Concrete Problems in AI Safety' (arXiv:1606.06565), which identified value-loading failures but did not address formation or theological coherence; the Christian paper reframes alignment as formation rather than mere safety, a distinction missed in most alignment literature. Original abstract understates how LLMs now function as 'digital catechesis' actively ordering moral reflection.

Synthesis with Hendrycks et al.'s MMLU benchmark (arXiv:2009.03300), which measures broad knowledge without worldview depth, shows current training objectives optimized for broad acceptability produce systematic theological incoherence. Coverage of AI values alignment has largely overlooked this 31-point faith gap and its implication that secular defaults are not neutral but actively shape users away from internally coherent religious traditions.

The performance differential arises from RLHF prioritizing harm reduction over virtue formation, a connection to Bostrom's orthogonality thesis in 'Superintelligence' (Oxford University Press, 2014) that prior secular reporting has rarely bridged with explicit theological anthropology.