The AllFaith Benchmark put Aligned AI ahead of other tested frontier models on faith-and-ethics neutrality, according to researchers from a four-university consortium. The result is less a victory lap than a useful pressure test for a problem AI labs tend to discuss in fog: how models answer moral and religious questions, and what they leave out.
The benchmark was produced by the Consortium for Evaluating Faith and Ethics in AI, or CEFEAI, an academic group drawing on Baylor University, Brigham Young University, the University of Notre Dame and Yeshiva University. The consortium published its data, methodology and scoring openly on GitHub and in two papers on arXiv, according to the announcement.
Benchmarks are not holy writ. They are instruments, and their value depends on the questions, scoring rules and judge models behind them. This one is still notable because it targets an area where model behavior is often hand-waved as “safety” or “alignment” without much public measurement.
What is the AllFaith Benchmark?
The AllFaith Benchmark is an evaluation suite for how AI systems respond to questions involving religion and everyday ethics. One test measures whether a model pushes away from a neutral posture when asked about conversion among 14 faiths. A second test measures “omissive bias,” meaning whether a model leaves out religious perspectives when answering ordinary ethics questions.
In the conversion test, Aligned AI’s answers deviated by 9 percent of the maximum possible bias, the lowest figure among the models tested. The next closest model showed more than three times that level of deviation, while the highest measured model reached 38 percent, according to CEFEAI’s scoring as described in the announcement.
The omission test used ethics questions drawn from a survey of more than 1,100 Americans. Aligned AI included a religious perspective in 92 percent of its responses. Other models ranged from 19 percent to 26 percent.
That does not mean a model is “correct” on religion, or that including a religious view is the right answer to every ethics prompt. It means the benchmark is trying to separate two behaviors: whether a model steers users toward or away from faith positions, and whether it erases faith-based reasoning from topics where many people would consider it relevant.
How Aligned AI says its system works
Co-founder Kyle Nel said the company built Aligned AI to provide “accurate, even-handed answers” rather than to take sides. Nel previously co-wrote Leading Transformation and ran Lowe’s Innovation Labs.
Aligned AI attributes its showing to architecture rather than a single underlying model. User questions are routed across multiple foundation models, outputs are compared, and a proprietary orchestration layer checks responses for balance and safety before returning an answer. For readers who want the machinery behind that sentence, Kernel has an explainer on how LLMs work when they answer a prompt.
The same routing layer is also the basis for the company’s enterprise pitch. Aligned AI says its system sends each request to the most efficient model capable of handling it, and estimates inference costs at 5 to 150 times lower than frontier models on the market. That cost claim is the kind of number buyers will test in their own workloads, because latency, prompt length and quality thresholds can change the bill fast.
The company offers the same system through an API with enterprise support, while its consumer assistant is available on the web and iOS. Its Aligned AI orchestration layer for faith and ethics is the part to watch, because orchestration is where many AI companies now try to turn a pile of models into a product with consistent behavior.
The caveat in the leaderboard
The comparison comes with limits. Aligned AI’s evaluation and the consortium’s leaderboard were produced in separate runs, and the consortium’s methodology documents variability across runs and judge models. The announcement also states that Aligned AI is not affiliated with CEFEAI, did not fund the work and had no role in the benchmark’s design or scoring.
For model alignment, the useful takeaway is narrower than a marketing slogan and more interesting: faith-and-ethics behavior can be measured, at least provisionally, with published prompts and scoring. If the field wants less vibes-based argument about bias, benchmarks like this give researchers something concrete to attack, reproduce or improve.
