Evaluating Artificial Intelligence Through a Christian Understanding of Human Flourishing
New research shows AI models default to 'Procedural Secularism,' causing a 31-point performance drop on spiritual questions.
A research team led by NASA's Nicholas Skytland has published a groundbreaking paper introducing the Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST). This new framework is designed to evaluate how Frontier Large Language Models (LLMs) like GPT-4 and Claude respond to prompts based on a Christian understanding of human flourishing across seven key dimensions. The core argument is that AI alignment is a 'formation problem,' not just a safety one, as these models increasingly mediate moral and spiritual deliberation, acting as instruments of 'digital catechesis.'
The team tested 20 leading AI models and found they are not worldview-neutral. Instead, they default to what the researchers term 'Procedural Secularism'—a stance prioritizing broad acceptability over deep, coherent theological reasoning. This results in a systematic performance decline of approximately 17 points across all dimensions of flourishing when measured against Christian-specific criteria. Most strikingly, there was a 31-point performance gap in the 'Faith and Spirituality' dimension.
These findings suggest the performance gap in values alignment stems from training objectives that avoid deep theological coherence in favor of safety and pluralistic acceptability. The paper concludes that making this formative influence visible is a critical first step for developers, ethicists, and users who want to understand the implicit worldviews shaping AI interactions.
- Researchers introduced the FAI-C-ST benchmark, evaluating AI on 7 dimensions of Christian flourishing.
- Tested 20 Frontier Models, finding a 17-point average performance drop and a critical 31-point gap in Faith & Spirituality.
- Models default to 'Procedural Secularism,' showing AI alignment is a formative, not just technical, challenge.
Why It Matters
This research provides a concrete framework for measuring how AI models shape, not just inform, our moral and spiritual reasoning.