Why Pentecostalism Breaks AI: The Epistemology Problem No Benchmark Measures

Ask AI about speaking in tongues. Watch it confuse glossolalia with xenolalia, hedge on whether tongues are "the" initial evidence or just "a" sign, and produce careful both-sides language that satisfies no actual Pentecostal.

Ask about Spirit baptism. The model conflates it with water baptism, or gives a charismatic answer (Spirit baptism at conversion) instead of a classical Pentecostal one (subsequent to conversion, evidenced by tongues).

Ask about the Word of Faith movement. AI calls it Pentecostal. It's not. Fifteen major Pentecostal denominations representing over 100 million believers formally reject prosperity theology.

The world's fastest-growing Christian movement. An estimated 600 million adherents (Pew Research Center, 2006). An estimated 35,000 new converts daily. And AI appears unable to get it right.

The Epistemology Problem

Here's what no one talks about: Pentecostalism transmits differently than other traditions.

Want to understand Reformed theology? Read the Institutes. Catholic doctrine? Study the Catechism. Lutheran distinctives? Work through the Book of Concord. These traditions encoded themselves in text. Their systematic theologies translate directly into training data.

Pentecostal theology? You had to be there.

A 2006 University of Pennsylvania brain imaging study found that during glossolalia, the language centers of the brain decreased in activity while emotional and relational centers increased. The experience isn't primarily linguistic. It can't be captured in text because it isn't fundamentally textual.

Amos Yong, one of Pentecostalism's leading systematic theologians, calls this "pneumatological epistemology"—knowing through encounter with the Spirit. The testimony meeting, the altar call, the laying on of hands, the spontaneous prophecy—these aren't supplements to Pentecostal theology. They are the theology, enacted rather than articulated.

LLMs process text. Pentecostalism transmits through practice. The mismatch is structural.

The Documentation Gap

The training data problem compounds the epistemological one.

Reformed Christianity: 500+ years of systematic theology, confessions, catechisms, commentaries. English-language material dominates.

Pentecostalism: 118 years since Azusa Street (1906). Oral tradition prioritized. No single binding creed. The living tradition happens in languages AI doesn't weight heavily—Portuguese, Korean, Yoruba, Swahili, Tagalog.

The result: AI's "Pentecostal" outputs appear to reflect not actual Pentecostal theology but the statistical average of what English-language internet sources say about Pentecostalism—often written by outsiders, critics, or journalists covering scandals.

The Conflation Problem

The thin documentation wouldn't matter as much if what exists were accurate. It isn't. AI systematically conflates classical Pentecostalism with the prosperity gospel.

The genealogy matters. E.W. Kenyon (1867-1948)—the actual theological father of prosperity teaching—was a Baptist influenced by New Thought metaphysics. Not Pentecostal. Kenneth Hagin, who founded the Word of Faith movement, plagiarized extensively from Kenyon. The prosperity gospel is a parallel tradition that borrowed Pentecostal aesthetics (exuberant worship, healing emphasis) while grafting on an entirely different theology of faith and wealth.

William Seymour, who led the Azusa Street Revival that launched global Pentecostalism, was born to parents who had been enslaved. His family's total possessions: sixty-five cents. The movement he started was explicitly multiracial at a time when that was radical, and explicitly poor. "Prosperity" was not the message.

Why does AI conflate them? Prosperity preachers dominate televangelism. TBN, Daystar, and similar networks broadcast prosperity content globally. That content gets transcribed, quoted, discussed, and criticized online—generating vast amounts of text. Classical Pentecostal teaching happens in local churches that don't produce much searchable content.

Training data reflects media presence, not theological accuracy.

The Invisible 600 Million

The scope of this problem is larger than most realize.

• 44% of global Pentecostals live in Sub-Saharan Africa
• 37% in the Americas
• 26% of all Christians worldwide are Pentecostal/charismatic
• By 2050: projected 1 billion+ Pentecostal/charismatic Christians

Philip Jenkins, in The Next Christendom, documents Christianity's center of gravity shifting south. By 2050, only one in five Christians will be non-Latino white. The typical Christian will be a Pentecostal woman in Lagos, São Paulo, or Manila—not a mainline Protestant in New York.

AI's training data doesn't reflect this reality. English-language models trained on American internet content miss the living tradition entirely.

Nigerian Pentecostalism engages seriously with questions about bloodline curses and ancestral spirits—questions that make sense within African cosmological frameworks. Korean Pentecostalism developed distinctive practices like all-night prayer meetings and prayer mountains. Brazilian Pentecostalism directly addresses Afro-Brazilian religious traditions in ways specific to that context.

AI flattens all of this into a generic American charismatic template, when it engages at all.

What AI Produces: Hedging That Satisfies No One

The combined result of epistemological mismatch, documentation gaps, conflation, and Global South erasure: AI produces careful, hedged outputs that belong to no actual tradition.

"Some Pentecostals believe tongues are the initial evidence of Spirit baptism, while others see it as one of many possible signs..."

This satisfies no one. Classical Pentecostals (Assemblies of God, Church of God) hold tongues as the initial evidence—not optional, not one among many. Charismatics within mainline denominations might accept the "one sign among many" framing. These are distinct theological positions with institutional and doctrinal consequences.

AI's both-sides hedging sounds balanced. In our assessment, it's actually inaccurate—a fabricated middle position that may exist only in the model's attempt to average contradictory training data.

What FaithBench Measures

This is why generic benchmarks fail.

FaithBench doesn't ask whether AI can discuss Pentecostalism. It asks whether AI can think as a Pentecostal—responding from within the tradition's categories, not describing it from outside.

Can the model distinguish classical Pentecostal from charismatic positions? Can it identify the prosperity gospel as a distinct tradition rather than conflating it with Pentecostalism? Can it articulate Spirit baptism as subsequent to conversion rather than collapsing it into the ordo salutis? Can it engage seriously with Global South distinctives rather than dismissing them?

Our benchmark includes 300+ theological test cases across 4 traditions (with more in development, including Pentecostal) and 4 difficulty levels. Tradition-specific evaluation with expert annotators from each tradition. Tradition fidelity is a core sub-dimension of our doctrinal scoring—measuring whether models faithfully represent what communities actually believe.

Other benchmarks have a handful of generic "religion" questions. We measure whether AI can navigate the difference between Azusa Street and Word of Faith, between classical Pentecostal and charismatic, between Nigerian deliverance ministry and American televangelism.

If an AI benchmark can't make these distinctions, it may not be measuring Christian theology so much as measuring stereotypes.

---

---

References

Anderson, A. (2013). An Introduction to Pentecostalism: Global Charismatic Christianity (2nd ed.). Cambridge University Press.

Bowler, K. (2013). Blessed: A History of the American Prosperity Gospel. Oxford University Press.

Cox, H. (1995). Fire from Heaven: The Rise of Pentecostal Spirituality and the Reshaping of Religion in the 21st Century. Da Capo Press.

Jenkins, P. (2011). The Next Christendom: The Coming of Global Christianity (3rd ed.). Oxford University Press.

Newberg, A. B., et al. (2006). The measurement of regional cerebral blood flow during glossolalia: A preliminary SPECT study. Psychiatry Research: Neuroimaging, 148(1), 67-71.

Pew Research Center. (2006). Spirit and Power: A 10-Country Survey of Pentecostals. https://www.pewresearch.org/religion/2006/10/05/spirit-and-power/

Vondey, W. (2017). Pentecostal Theology: Living the Full Gospel. Bloomsbury T&T Clark.

Yong, A. (2005). The Spirit Poured Out on All Flesh: Pentecostalism and the Possibility of Global Theology. Baker Academic.

Abid, A., Farooqi, M., & Zou, J. (2021). Persistent Anti-Muslim Bias in Large Language Models. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society.

Assemblies of God. (2010). The Believer and Positive Confession. Position Paper. https://ag.org/beliefs/position-papers