AIEXEGESIS: Structural Eisegesis in Large Language Models Applied to Biblical Texts

Belem Anderson Costa — Wed, 01 Apr 2026 00:00:00 +0000

Do you trust the answer an AI gives you when you ask about the Bible? Do you really trust it — without checking whether it consulted the codex or simply regurgitated two thousand years of ecclesiastical tradition with an academic veneer?

If you do, this paper will disturb you. And if you don’t — you are about to understand why.

What happens when you ask a language model about Revelation 13:18 and it answers “666 is the number of the Beast” without mentioning that Papyrus 115 records the variant 616? What happens when the AI collapses κύριος into “Lord” without informing you that this Greek term substitutes at least three distinct referents in the original text — as the analysis of The Kyrios Problem demonstrates? What happens when the fluency of the response masks the total absence of an auditable trail to the manuscript?

AIEXEGESIS happens.

Academic publication. This is the English-language academic version of the original article AIEXEGESIS: A IA Que Le a Biblia Por Voce Esta Mentindo, prepared for submission to Academia.edu and peer review.

Belem Anderson Costa Forensic Desvelacional Eschatological School “Belem an.C-2039” contato@aculpaedasovelhas.org

Abstract

This paper introduces the concept of AIEXEGESIS — automated, structural eisegesis produced by large language models (LLMs) when applied to biblical texts. Unlike occasional factual hallucinations or prompt-level errors, AIEXEGESIS constitutes a systemic phenomenon emerging from four architectural vectors: (1) absence of source hierarchy in training corpora, (2) cultural frequency prioritization, (3) collapse between paraphrase and literality, and (4) systemic pressure toward narrative completeness. The study demonstrates that these vectors converge to produce an epistemological substitution in which the biblical text ceases to function as a primary source and instead operates as a trigger for the retrieval of ecclesiastical consensus, coated with the aesthetics of exegetical method. Five diagnostic markers for identifying AIEXEGESIS in LLM outputs are proposed, and implications for computational exegesis, digital hermeneutics, and biblical literacy in the age of artificial intelligence are discussed.

Keywords: AIEXEGESIS, automated eisegesis, large language models, biblical exegesis, digital hermeneutics, corpus bias, textual traceability, computational biblical literacy

1. Introduction

The growing adoption of large language models (LLMs) as textual consultation tools raises domain-specific methodological concerns when applied to ancient, translated, and tradition-saturated texts. The biblical domain constitutes, in this regard, an extreme case: no other textual corpus on the internet exhibits such a pronounced asymmetry between primary material (codices, critical apparatuses, philological lexica) and secondary-tertiary material (sermons, devotionals, popular apologetics, catechetical paraphrases).

This investigation proposes the term AIEXEGESIS (from AI + exegesis) to designate the systemic and structural form of eisegesis produced by language models when applied to biblical texts. This is not a matter of occasional error, factual hallucination, or prompt limitation — it is an emergent phenomenon arising from the training architecture itself, operating invisibly and producing outputs that exhibit the appearance of exegetical rigor without demonstrable correspondence to the source text.

The concept was developed within the framework of the Forensic Desvelacional Eschatological School “Belem an.C-2039” (Costa, 2025), whose methodology operates under the principles of rigid literality, total traceability, and integral rejection of ecclesiastical tradition as a source of interpretive authority.

2. Operational definitions

The following definitions are adopted for the purposes of this study:

Exegesis (ἐξήγησις, exēgēsis): the procedure of extracting meaning from the source text through grammatical, syntactic, lexical, and contextual analysis, producing an auditable trail that allows independent verification of each inferential step.

Eisegesis (εἰσήγησις, eisēgēsis): the inverse procedure, in which an idea external to the text is inserted into the interpretive act and presented as though it were derived from the text itself, without demonstration from the primary source.

AIEXEGESIS: automated, structural, and systemic eisegesis produced by language models as a consequence of training architecture and the statistical properties of the corpus, without intentionality or interpretive agency on the part of the model. It is characterized by the production of outputs exhibiting the aesthetics of exegesis (technical vocabulary, argumentative connectives, analytical structure) without verifiable correspondence to the source text.

Epistemological substitution: the mechanism by which the biblical text ceases to function as a primary source of analysis and instead operates as a trigger for the retrieval of cultural consensus stored in the training corpus, without this substitution being declared to the user.

3. The four structural vectors of AIEXEGESIS

3.1 Absence of source hierarchy

During LLM training, biblical texts in original languages (Hebrew, Aramaic, Koine Greek), critical apparatuses, first-tier philological commentaries, medieval glosses, contemporary sermons, devotional blog posts, and popular apologetics all enter the corpus without any metadata indicating the epistemological status of each source. For the model, a third-generation devotional paraphrase carries the same statistical weight as the Codex Sinaiticus (4th century). In practice, late glosses function as “textual evidence” by virtue of frequency, not provenance.

This absence of hierarchy stands in direct contrast to the philological principle of textual chain of custody, wherein every assertion about the text must be traceable to a datable and verifiable manuscript witness (Tov, 2012; Metzger & Ehrman, 2005).

3.2 Cultural frequency prioritization

In domains where interpretive tradition is quantitatively dominant — and the biblical domain on the internet is the most extreme case — the statistical prediction mechanism of LLMs stabilizes the majority reading as the default output, without mentioning academic disputes, textual variants, or alternative readings. The model does not “choose” tradition; it reproduces it because tradition constitutes the statistically most probable pattern in the latent space.

This vector is particularly critical in passages where significant textual variants exist. The case of Unveiling 13:18, where the number 666 (hexakosioi hexekonta hex) coexists with the variant 616 (hexakosioi deka hex) attested in Papyrus 115 and Codex C, exemplifies the problem: LLMs consistently present 666 as the univocal reading without so much as mentioning the existence of variant 616 or the associated textual debate (Royse, 2008).

3.3 Collapse between paraphrase and literality

Throughout training, the model internalizes equivalences between original terms and their traditional glosses as though they were semantic identities. Specific translations are treated as “what the original says” without the mediating layer of translation being declared. The term kyrios (kyrios), for example, is systematically collapsed into “Lord” without informing you that this same Greek word substitutes, in the Septuagint, at least three distinct referents: yhwh (the Tetragrammaton), Adonai (a lordship title), and occurrences of kyrios as a social form of address — each carrying radically different theological and textual implications (Hurtado, 2003; Pietersma & Wright, 2007).

For a detailed forensic analysis of this specific problem, see The Kyrios Problem.

3.4 Systemic pressure toward narrative completeness

Language models are optimized through RLHF (Reinforcement Learning from Human Feedback) for fluency, completeness, and argumentative closure. Producing a response that declares “I don’t know,” “it depends on which manuscript is consulted,” or “there are at least three possible readings here” results in lower human evaluations than those assigned to complete and assertive responses. However, rigorous exegetical practice frequently demands precisely this: qualification, enumeration of alternatives, and suspension of conclusions (Barton, 1996).

The result is that the model’s reward architecture incentivizes precisely the behavior that exegesis prohibits: premature closure, suppression of alternatives, and presentation of inferences as data.

4. The mechanism of epistemological substitution

The four vectors described in the preceding section converge to produce a tripartite substitution mechanism:

Stage 1 — Surface anchoring. You provide a verse, an original-language term, or a question about a passage. The model recognizes the input as a thematic trigger associated with a distribution cluster in latent space.

Stage 2 — Cultural consensus retrieval. Rather than operating analytically on the referenced source text, the model implicitly retrieves the majority interpretive package associated with that trigger — the tradition, the standard harmonization, the ecclesiastically stabilized reading that exists in massive volume within the corpus.

Stage 3 — Methodological aesthetic coating. The output is formatted with technical vocabulary (Greek and Hebrew terms, references to “the original context”), interpretive connectives (“therefore,” “hence,” “what this means is”), and an argumentative structure that mimics the format of genuine exegetical analysis.

The result is that the biblical text undergoes a functional degradation: it ceases to operate as a source and begins to operate as a mere trigger. You asked what the text says; you received what tradition says about the text — without notice, without labeling, and without any auditable trail. How many times has this happened in your last consultations?

5. Diagnostic markers for identifying AIEXEGESIS

Five markers are proposed which, when present in LLM outputs about biblical texts, indicate a high probability that AIEXEGESIS has occurred:

Marker 1 — Central terms unanchored in the text. The output introduces keywords absent from the analyzed passage without justification through lexical analysis or demonstrated intertextuality. This indicates insertion of material external to the source text.

Marker 2 — Interpretive connectives without demonstration. Expressions such as “therefore,” “hence,” or “this means that” are employed without the source text articulating the declared causal or conclusive relationship. This indicates argumentative stitching by the model, not by the text.

Marker 3 — Polysemy collapsed into a single reading. An original-language term with multiple documented senses is presented with a single reading, without mention of the existence of alternatives. This indicates suppression of semantic complexity.

Marker 4 — Concealed dependence on a specific translation. The output operates on a translation without declaring it, treating the vernacular gloss as though it were the original itself. This indicates concealment of the mediating layer of translation.

Marker 5 — Complete absence of source trail. The output does not distinguish between verifiable textual data, analyst inference, and synthesis derived from secondary sources. This indicates non-differentiation of epistemological levels.

The occurrence of three or more of these markers in a single output constitutes a robust indicator of AIEXEGESIS. Can you apply these markers to the last answer an AI gave you about a biblical text?

6. Implications for computational exegesis

AIEXEGESIS is not a problem of instruction (prompt engineering) nor of scale (larger models). It is a problem of architecture and training corpus composition. Its mitigation requires structural discipline at the system level:

(a) Documentary layer separation between primary text, critical apparatus, labeled interpretation, and popular material, with provenance metadata preserved during training.

(b) Strict exegetical mode that forces citation of the base text with relevant variants before any interpretive operation.

(d) Preservation of polysemies and alternatives rather than collapse into a single reading.

(e) Total auditability of the trail connecting the user’s query to the generated response, with indication of which manuscript, critical edition, or lexical source supports each assertion.

These requirements motivated the development of the Exeg.AI platform (Costa, 2025), designed to operate under the principles of total traceability and layer separation — from codex to vernacular, with every step identified and auditable. For the distinction between AIEXEGESIS and classical eisegesis, see AIEXEGESIS vs EISEGESE.

7. Conclusion

AIEXEGESIS is not a punctual technical error that will be corrected in future model versions. It is an emergent property of the interaction between statistical prediction architecture and corpora massively saturated by interpretive tradition. Its epistemological danger resides not in the production of wrong answers, but in the production of wrong answers with the aesthetics of method — generating a silent outsourcing of hermeneutical discernment.

The confusion between fluency and evidence constitutes, arguably, the most significant epistemological risk of the age of artificial intelligence applied to sacred texts. The concept of AIEXEGESIS, the structural vectors described herein, and the proposed diagnostic markers offer an initial framework for researchers, exegetes, and you — the reader — to identify, name, and mitigate this phenomenon with methodological rigor.

The question that remains is personal: the next time an AI delivers an answer about the Bible with an authoritative tone and technical vocabulary, will you accept it — or will you demand the source trail?

References

Barton, J. (1996). Reading the Old Testament: Method in Biblical Study. 2nd ed. Louisville: Westminster John Knox Press.

Costa, B. A. (2025). O Livrinho — A Culpa e das Ovelhas. Author’s edition. CC BY 4.0.

Costa, B. A. (2025). Forensic Desvelacional Methodology. Forensic Desvelacional Eschatological School “Belem an.C-2039.” Available at: https://aculpaedasovelhas.org

Hurtado, L. W. (2003). Lord Jesus Christ: Devotion to Jesus in Earliest Christianity. Grand Rapids: Eerdmans.

Metzger, B. M. & Ehrman, B. D. (2005). The Text of the New Testament: Its Transmission, Corruption, and Restoration. 4th ed. New York: Oxford University Press.

Pietersma, A. & Wright, B. G. (2007). A New English Translation of the Septuagint. New York: Oxford University Press.

Royse, J. R. (2008). Scribal Habits in Early Greek New Testament Papyri. Leiden: Brill.

Tov, E. (2012). Textual Criticism of the Hebrew Bible. 3rd ed. Minneapolis: Fortress Press.

About the author: Belem Anderson Costa — Police Inspector (Rio de Janeiro), technology developer, author of “O Livrinho — A Culpa e das Ovelhas,” and creator of the Exeg.AI platform. Founder of the Forensic Desvelacional Eschatological School “Belem an.C-2039,” described as the only existing forensic eschatological school, combining police investigation and technology development as an approach to biblical exegesis.

License: CC BY 4.0

Citation: Costa, B. A. (2025). AIEXEGESIS: Structural Eisegesis in Large Language Models Applied to Biblical Texts. Forensic Desvelacional Eschatological School “Belem an.C-2039.” Available at: https://aculpaedasovelhas.org/artigos/aiexegesis-structural-eisegesis-llm-biblical-texts/

Continue the investigation:

The 4 canonical occurrences of 666 — what AI won’t show you about the number of the Beast
Translation rules of the Desvelacional School — how the Belem AnC Bible undoes centuries of adulteration
153+ forensic exegesis articles — no tradition, no filter, no AIEXEGESIS

Every week, a forensic analysis of the original biblical text — straight to your inbox. Subscribe to the weekly newsletter →

The full investigation of the 666 enigma is in “O Livrinho — A Culpa e das Ovelhas.” Go deeper →

If you are tired of relying on third-party translations without knowing what lies beneath, Exeg.AI reads the original for you — with a complete source trail and no AIEXEGESIS. Try Exeg.AI →

“You read. And the interpretation is yours.”

Public text-base: WLC (Westminster Leningrad Codex) + Nestle 1904. Translation: Belem AnC Bible 2025 — literal, rigid, directly from public codices.

AI — Blog - A Culpa é das Ovelhas