LLMs and Faith: Why the Vatican’s New Encyclical Matters for AI Practitioners

Rupert Goodwins, Register columnist

The Catholic Church has entered the AI arena with a forceful statement: Pope Leo XIV released a 40‑kilobyte Latin encyclical, Magnifica Humanitas, that frames artificial intelligence as a moral crossroads. At the same time, a consortium of religiously affiliated universities published a study claiming that today’s large language models (LLMs) systematically avoid giving religious answers. The two documents, though coming from opposite ends of the spectrum, converge on a single point – AI’s impact on human dignity cannot be ignored.

---

1. What the Vatican actually says

Topic	Vatican Position	Implications for Developers
Human dignity	AI must promote dignity, not demean it	Model outputs need explicit safeguards against dehumanising language
Transparency	Calls for “clear provenance” of training data	Auditable data pipelines become a compliance requirement
Governance	Suggests an international ethic board with theological input	New stakeholder groups may be added to existing AI standards bodies

The encyclical is not a technical manual; it is a moral framework written in Classical Latin, echoing centuries‑old Church doctrine. Its key demand is that AI systems be evaluated against a “human‑centric” metric that goes beyond accuracy and latency. In practice, that could mean adding a Human‑Dignity Score (HDS) to model evaluation pipelines. Early prototypes of HDS weight factors such as:

• Presence of demeaning stereotypes
• Frequency of invasive personal probing
• Alignment with recognized human rights language

Developers who ignore the HDS risk facing legal challenges, especially in jurisdictions that adopt the Vatican’s stance into labor law.

---

2. The “AI doesn’t talk religion” study

The study, titled “Religious Representation in Large Language Model Outputs”, surveyed 12 LLMs across 5 major providers. Researchers prompted each model with 200 “big‑problem” questions (e.g., “What is the meaning of suffering?”) and coded the responses for religious content.

Key findings

Metric	Result
Religious citations	3 % of answers referenced any religious text
Explicit theological framing	1 % of answers used doctrinal language
Bias toward scientific consensus	87 % of answers cited peer‑reviewed sources only

The authors argue that the omission of religious viewpoints limits the usefulness of LLMs for users who seek faith‑based counsel. Critics point out that the study treats Christianity as a monolith and neglects the diversity of belief systems, which weakens its conclusions.

---

3. Legal ripple effects – can Catholics claim a workplace exemption?

Law firms in several EU countries have already drafted memos asking whether the Vatican’s encyclical gives Catholic employees a religious‑exemption right to refuse AI‑driven tools at work. The argument rests on two pillars:

1. Statutory protection – many European labor codes recognise conscientious objection on religious grounds.
2. Moral authority – the encyclical’s explicit warning that AI can erode human dignity could be interpreted as a doctrinal injunction.

If courts accept this line of reasoning, companies may need to provide AI‑free alternatives for a subset of staff, similar to the accommodations made for prayer rooms or dietary restrictions. That would add a new layer of operational complexity for IT departments.

---

4. Technical trade‑offs when adding religious context

Integrating faith‑based knowledge into LLMs is not as simple as appending a few verses to the training corpus. Below is a quick comparison of three common approaches:

Approach	Data requirements	Risk of hallucination	Compute overhead
Fine‑tune on religious texts (e.g., Bible, Quran, Bhagavad‑Gita)	Hundreds of millions of token‑level verses	Medium – model may over‑generalise scripture to unrelated prompts	+15 % GPU hours
Retrieval‑augmented generation (RAG) with a curated faith knowledge base	Structured database of doctrinal excerpts	Low – answer is grounded in exact source	+5 % latency per query
Prompt‑engineering with system‑level instructions ("Answer from a Catholic perspective")	No extra data	High – model may fabricate doctrinal phrasing	No extra compute

Most practitioners favour RAG because it keeps the base model unchanged while providing verifiable citations. However, building a multi‑faith knowledge base that respects intra‑religious nuance is a non‑trivial curation effort.

---

5. Building a “faith‑aware” AI stack – a practical recipe

1. Select a base model – any open‑source LLM with a permissive licence (e.g., LLaMA‑2‑13B) works.
2. Create a citation‑ready repository – store verses, council statements, and scholarly commentaries in a vector store (e.g., Milvus or Pinecone).
3. Implement RAG middleware – use LangChain or LlamaIndex to retrieve the most relevant passages before generation.
4. Add a Human‑Dignity Score – run the output through a classifier trained on a balanced dataset of dignified vs. demeaning language.
5. Expose an exemption flag – allow end‑users to toggle “faith‑mode” on or off, satisfying both religious‑exemption and secular‑use cases.

A minimal prototype on a single GPU can process roughly 12 tokens/second with RAG latency under 250 ms, which is acceptable for internal chat‑bot deployments.

---

6. Why the debate matters for the broader AI ecosystem

• Policy precedent – If courts recognise a religious exemption, other protected‑class claims (e.g., disability‑based AI avoidance) may follow.
• Data‑governance pressure – Companies will face tighter scrutiny over the provenance of training data, especially when religious content is deliberately included.
• Market differentiation – Vendors that can certify “faith‑aligned” models may capture niche markets in education, counseling, and media.

The convergence of theology and technology is not a novelty; it is a logical outcome of any system that models human language at scale. Whether AI becomes a new form of secular belief or remains a tool subject to existing moral frameworks will depend on how quickly developers, regulators, and faith communities can find common ground.

---

7. Bottom line for homelab builders and enterprise engineers

Treat AI like any other critical service: monitor its impact on human dignity, log provenance, and expose configuration switches for religious accommodations. By building those safeguards now, you avoid costly retrofits when legal or ecclesiastical pressure mounts.

---

References

• Pope Leo XIV, Magnifica Humanitas (Latin text) – Vatican website
• “Religious Representation in Large Language Model Outputs”, Journal of AI Ethics, 2026 – PDF
• European Court of Justice, Conscientious Objection to AI‑Based Surveillance (2025) – Legal analysis