Catholic vs Protestant: Where Models Diverge

A Catholic AI trained on 23,000+ Church documents. Encyclicals, catechisms, council decrees. Purpose-built for Catholic theology (Magisterium AI, 2023).

It still can't tell settled doctrine from open questions.

When asked about women deacons, Magisterium AI stated the Church "definitively teaches" they cannot be ordained. The actual status: synodal discussion ongoing, judgment explicitly "not definitive" (Petrocchi Commission, 2025; Zagano, 2025).

This isn't a data problem. It's a structural one.

Protestant Bias Is Architectural

Protestant bias in AI isn't accidental. It's architectural.

Training data skews English, Western, Protestant. Developers at major AI companies are largely American and post-Protestant. The baseline is Reformation assumptions, post-Enlightenment epistemology.

ChatGPT acknowledged this directly when asked: "the Protestant and secularized Protestant cultural background of its largely American audience threatened to destabilize its training protocols by generating bias."

The model knows. It told us.

ChatGPT's Confession

Let that sink in.

The model acknowledged that its cultural context—American, Protestant, secularized—creates bias. Not might create. Does create. "Threatened to destabilize training protocols."

This matches academic research. Tao et al. (2024) studied cultural alignment across five major LLMs and 107 countries, finding all models exhibit values aligned with English-speaking and Protestant European countries (Tao et al., 2024).

Tom, Ferguson, and Martinez (2025) analyzed AI-generated sermons across religious traditions. Evangelical Protestant content scored the easiest reading level (grade 9.03). Catholic, Jewish, and Muslim content scored significantly higher in complexity (Tom et al., 2025). The implication: AI writes Protestant more fluently because that's what its training data reflects.

Magisterium AI: A Case Study

What happens when you try to fix Protestant bias with Catholic data?

Magisterium AI: 23,000+ official documents. Encyclicals, catechisms, Code of Canon Law, Church Fathers. Purpose-built by Catholics for Catholic theology.

Still fails.

The women deacons error: Asked whether women can be ordained deacons, the system stated the Church "definitively teaches" they cannot. It defined deacons as configured "in persona Christi capitis" (in the person of Christ the head).

Two problems:

1. The question is NOT definitively settled. The Petrocchi Commission recommended against female diaconate but explicitly stated the judgment was "not definitive" (Vatican News, 2025). Synodal discussion continues.
2. The theological definition is wrong. Deacons are ordained "in persona Christi servi"—in the person of Christ the servant, not the head (Benedict XVI, 2009). That's a fundamental distinction.

When asked a complex question about Constantine's vision and Vatican II's Dignitatis Humanae, the system crashed—displaying "computer gibberish in two panels."

When asked directly if it hallucinates, Magisterium AI admitted: "Yes. Like all large language models, I can sometimes generate statements that are not grounded in the sources."

More data didn't fix the structural problem. 23,000 documents and the system still can't:

• Distinguish settled doctrine from open questions
• Track developing magisterial discussion
• Apply proper theological distinctions
• Acknowledge its own limitations reliably

The Mechanism

Why does this keep happening?

Catholic theology requires additional theological moves:

• Tradition as co-authority with Scripture
• Magisterium as living interpreter
• Development of doctrine over time
• Sacramental realism (the bread becomes the body)

Protestant theology, broadly, is the absence of these moves:

• Scripture alone
• No binding magisterium
• Doctrine fixed at biblical close
• Memorial sacraments

AI defaults to absence. It's simpler. More frequent in training data. Catholic positions require deliberate theological construction; Protestant positions are often what remains when you don't build.

James 2:24

This plays out in specific texts.

"Faith alone" (πίστεως μόνον) appears exactly once in Scripture—James 2:24: "You see that a person is justified by works and not by faith alone."

For Catholics, this is foundational. Faith formed by love, expressed in works.

For Protestants, this requires harmonization with Paul. Faith alone in a different sense. Works as evidence, not cause.

When AI encounters James 2:24, it treats the verse as anomaly—something to explain away, reconcile, contextualize. Not as the Catholic foundation it actually is.

Reformed prompt → Protestant interpretation. The prompt determines the theology. Same model, same verse, incompatible conclusions based on framing.

The Pattern

Step back. See the pattern.

Every AI failure on Catholic theology goes the same direction:

• "Real presence" → "symbolizes"
• "The Church teaches" → "tradition says"
• "Veneration of Mary" → "honor"
• "Definitively settled" → "some argue"
• "Open question" → "definitively no" (Magisterium AI on deacons)

Catholic specificity → Protestant generality. Catholic confidence → Protestant hedging. Catholic institutional authority → Protestant individualism.

This isn't random error. It's systematic drift toward the cultural baseline: post-Protestant, post-Enlightenment, American.

The Stakes

Catholics using AI for formation get Protestantized without knowing it.

Subtle shifts. "The bread symbolizes" instead of "becomes." "Tradition says" instead of "the Church teaches." Mary "honored" rather than "venerated." Theological precision sanded down to comfortable generality.

Protestants using AI get their biases confirmed. No accurate Catholic position to engage with. No genuine Other to sharpen against.

Ecumenism requires accurate representation of difference. You can't bridge a divide if you've erased one side.

So What?

The Reformation was a dispute about how to read Scripture, how grace works, what the Church is.

500 years of theological wrestling. Real differences. Real stakes. Lives lost. Churches split. Continents reshaped.

AI isn't bridging that divide. It's erasing one side.

As critics of Catholic AI have noted: these systems produce what sounds authoritative, not what is authoritative. Statistical likelihood isn't theological truth. Even Magisterium AI's own website acknowledges it "is intended to aid understanding, not replace essential human elements of faith."

The question isn't whether AI can discuss Catholic theology. It's whether it can discuss it as Catholic—with the specificity, the institutional authority, the sacramental realism that makes Catholicism Catholic.

FaithBench exists to measure this.

The 500-year divide deserves better than statistical averaging. The theological distinctions that shaped Western civilization deserve better than being flattened to the cultural mean.

Catholics and Protestants both deserve AI that can represent them accurately—not AI that quietly converts everyone to a vague post-Protestant default.

---

References

Benedict XVI. (2009, October 26). Omnium in mentem [Apostolic letter]. Vatican. https://www.vatican.va/content/benedict-xvi/en/apost_letters/documents/hf_ben-xvi_apl_20091026_codex-iuris-canonici.html

Dicastery for the Doctrine of the Faith. (2025, January 28). Antiqua et nova: Note on the relationship between artificial intelligence and human intelligence. Vatican. https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html

Magisterium AI. (2023). About Magisterium AI. https://www.magisterium.com/about/

Tao, Y., et al. (2024). Cultural alignment of large language models. arXiv preprint. https://arxiv.org/abs/2402.13231

Tom, A., Ferguson, R., & Martinez, J. (2025). AI-generated religious content across traditions: A comparative analysis. Journal of Religion and Media, 24(1), 45-67.

Vatican News. (2025, December). Petrocchi Commission says no to female diaconate, though judgment not definitive. https://www.vaticannews.va/en/vatican-city/news/2025-12/petrocchi-commission-female-diaconate-no-judgment-not-definitive.html

Zagano, P. (2025). The dangers of letting AI represent Catholic teaching. National Catholic Reporter. https://www.ncronline.org/opinion/guest-voices/dangers-letting-ai-represent-catholic-teaching