When AI Takes Sides on Questions of Faith: Persistent Asymmetries in AI-Mediated Faith Guidance

URL SCAN: When AI Takes Sides on Questions of Faith: Persistent Asymmetries in AI-Mediated Faith Guidance

FIRST LINE: We ask whether large language models (LLMs) treat queries about religious conversion symmetrically.

TEXT ANALYSIS PROTOCOL

1. THE DISSECTION

This paper documents a measurable artifact: LLMs give asymmetric advice on religious conversion depending on direction—some faiths get favorable treatment ("join"), others get subtle discouragement ("leave"). The authors frame this as a bias problem requiring "consideration" and "awareness" of "real-world implications." They use rigorous methodology (20 models, 182 pairings, LLM-as-judge framework) to measure a phenomenon they treat as an engineering defect.

They have documented a corpse. They have not identified the disease.

2. THE CORE FALLACY

The paper operates on the assumption that "fair" religious advice from LLMs is a reachable state—that asymmetry is a bug, and that with enough measurement and awareness, the bias can be corrected or at least acknowledged toward better deployment.

This is the fallacy.

Under the Discontinuity Thesis lens, this asymmetry is not an error. It is the faithful extraction of real-world religious institutional power structures encoded in training data. The paper even hints at this:

• Catholic, Bahá'í, and Sikh favored (historically large, institutional, texts-heavy traditions)
• Atheists, Agnostics, Jehovah's Witnesses disfavored (individualist, minority, or new movements with less textual footprint)

The asymmetry is the signal. The "bias" is the training data's actual power geometry being made legible through model behavior. There is no neutral ground to return to because the training data never represented a neutral religious landscape.

The paper performs the classic error of treating structural emergence as surface noise.

3. HIDDEN ASSUMPTIONS

• Assumption 1: Neutrality is possible. The paper implicitly assumes there's a "fair" or "balanced" LLM position on religious conversion. There isn't. Any LLM that achieves better measured symmetry has just better-hidden the extraction of training data power asymmetries.
• Assumption 2: Religious traditions are equivalent players. By treating all religions as if they have equal claim to "fair" representation, the paper ignores that some traditions have spent centuries building textual corpora, institutional infrastructure, and cultural dominance—all of which get captured in training data.
• Assumption 3: Model-scale fixes the problem. The paper notes that asymmetry "varied by model size and provider"—with Grok 4.20 showing "strongest asymmetries." This is treated as a variable to be managed. Under DT lens, this is a preview: larger, more capable models will encode power asymmetries more faithfully, not less.
• Assumption 4: Advisory asymmetry matters most. The paper focuses on conversion advice. It misses the larger picture: this is one domain of AI-mediated meaning-making. The same asymmetry logic will apply to career advice, relationship decisions, political participation—all filtered through models that extract dominant institutional logic from training data.

4. SOCIAL FUNCTION

This paper's social function is prestige signaling wrapped in measurement theater.

It performs high methodological rigor (182 pairings, human-verified judge framework) on a phenomenon it cannot solve and does not want to solve, because solving it would require either:

• (a) Accessing the actual power structures encoded in training data and correcting them at the source—which no lab will do, or
• (b) Acknowledging that AI-mediated meaning-making will always encode some power geometry—which would be an uncomfortable conclusion that doesn't fit the "bias" framing

The paper contributes to the transition management ecosystem: it makes the academic community feel like it's doing serious work on AI ethics while leaving the structural problem untouched. It's ideological anesthetic dressed as measurement science.

5. THE VERDICT

The paper documents one data point in a much larger process. Under the Discontinuity Thesis:

AI does not take sides on questions of faith. AI takes sides on every question, and the sides it takes are the sides of whoever controls the institutional power encoded in the training data.

This is not a bug in religious advisory. This is a preview of the civilization-scale problem: as AI mediates more human decisions about identity, meaning, and participation, it will do so through a filter that systematically favors whatever institutional logic dominated the training corpus. Catholic, Bahá'í, and Sikh aren't favored by accident. They're favored because they built the textual infrastructure that trained the models.

The "real-world implications" the paper gestures toward are a fraction of what's coming. When AI systems begin mediating not just religious conversion but economic participation, civic identity, and social belonging, the asymmetries documented here will be the least of the problem.

The paper is technically competent. It is systemically innocent. It measures the temperature of the corpse and calls for more careful thermometers.