Can
Generative AI Summarize Sacred or Biased Texts Fairly? Ethical
Implications
When a large language model is asked to summarise the Quran, the
question it is actually being asked is not a simple retrieval task. It is a
request to compress centuries of scholarship, interpretation, spiritual
practice, and lived community meaning into a few paragraphs generated by a
system trained primarily on text produced by people who were not Muslim,
writing in a language that is not Arabic, for audiences that are not the
communities for whom the Quran is sacred. The same challenge applies to the
Torah, the Bible, the Bhagavad Gita, the Pali Canon, and every other sacred
text that AI systems are now being asked to process.
The difficulty is not merely technical. It is epistemic. Sacred
texts are not just repositories of information; they are living documents
whose meaning is inseparable from the communities that interpret them, the
traditions of commentary that surround them, and the contexts of use, prayer,
study, ritual, legal reasoning, that give individual passages their
significance. A generative AI system that reduces a sacred text to a summary
is not summarising that text in any meaningful sense. It is producing a
statistical compression of surface features that may be accurate at the level
of factual content and profoundly misleading at the level of
meaning.
What the Research Shows
The empirical evidence for AI bias in religious text summarisation
is accumulating. A study
published in Scientific Reports tested multiple large language
models on their handling of religious content and found measurable cognitive
bias in how models interpreted and summarised material from different faith
traditions. The bias pattern was consistent: models performed most accurately
on content from traditions that are well represented in English-language
training data, predominantly Western Christian sources, and less accurately
on content from traditions with smaller English-language digital footprints
or whose canonical languages are not English.
The mechanism is not mysterious. Language models learn to generate
text by predicting what tokens are likely to follow given what precedes them,
based on statistical patterns in training data. When the training data
includes substantially more commentary, analysis, and contextualisation of
one religious tradition than others, the model’s representation of the
under-represented traditions is less rich, less nuanced, and more prone to
defaulting to majority-tradition frameworks when interpreting ambiguous
material.
Pangeanic’s
analysis of AI religious text translation identifies the specific
problem of theological subtext in translation contexts. Sacred texts
frequently employ language that carries multiple layers of meaning
simultaneously, with specific words resonating differently depending on the
interpretive tradition of the reader. AI translation systems optimised for
semantic accuracy at the surface level may consistently fail to convey these
deeper resonances, producing outputs that are technically correct and
spiritually inadequate.
Faith-Based AI and the Doctrinal Echo Chamber
Risk
The development of AI systems specifically designed for religious
communities raises a different set of concerns. Magisterium AI, a Catholic AI
assistant launched in 2025 and reported
on by the Washington Post, is designed to provide answers grounded
in official Catholic teaching. The system represents an attempt to solve the
bias problem by deliberately aligning the AI with a specific interpretive
tradition rather than claiming to provide neutral
summarisation.
The appeal is understandable: a Catholic user asking about the
Church’s position on a moral question benefits from an AI that understands
the doctrinal context. But the risk is the inverse of the bias problem. A
system that provides only the official position of one interpretive
tradition, without acknowledging the diversity of opinion within that
tradition or the perspectives of those who dissent from it, may reinforce a
uniformity of religious expression that does not accurately reflect the
community it claims to serve.
Both problems, the secular AI’s tendency to compress religious
meaning toward majority frameworks and the faith-aligned AI’s tendency to
enforce doctrinal uniformity, point to the same underlying difficulty: the
relationship between a text and its community of interpretation is not
something an AI system can be given as a parameter. It is a living
relationship that evolves continuously and that no system trained on a static
dataset can fully represent.
The Political Text Problem
The challenges of fair summarisation are not limited to religious
material. Politically charged texts, historical documents produced from
partisan perspectives, and ideologically shaped narratives present analogous
difficulties. An AI system asked to summarise a political manifesto, a
colonial-era historical account, or a document produced by an advocacy
organisation must navigate the same tension between surface accuracy and
deeper meaning that characterises sacred text processing.
The arXiv paper Modeling the
Sacred argues for what it calls cultural positionality in AI
systems: the explicit acknowledgment by a system of its interpretive
standpoint rather than a claim to neutral summarisation. The argument extends
naturally to politically and historically charged texts: a system that
acknowledges the perspective from which it is approaching a document is more
honest and ultimately more useful than one that presents its output as
objective.
As explored in AI
and Emotional Manipulation, the framing of AI outputs has
significant effects on how audiences receive and process the information they
contain. A summary that strips out the emotional and ideological texture of a
charged text is not a neutral product. It is a particular kind of reframing
that may serve some purposes and mislead others.
Educational Implications
The use of AI summarisation tools in educational contexts involving
religious or politically sensitive texts raises specific concerns about the
development of critical reading skills. Students who encounter sacred or
charged texts through AI summaries rather than direct engagement with the
original material are receiving a mediated version that may not accurately
represent the complexity of what they are being asked to
study.
As discussed in Teaching
Tomorrow, the relationship between AI tools and educational
development requires careful design. In the specific context of religious and
cultural education, where the goal is not just comprehension but appreciation
of diversity and complexity, AI summarisation may actively work against
educational objectives if deployed without explicit framing about its
limitations.
The practical implication is not that AI tools should be excluded
from religious or cultural education but that their use should be accompanied
by explicit instruction in the limitations of AI summarisation for complex,
culturally embedded texts. Students who understand why an AI summary of the
Quran is necessarily incomplete are better equipped to engage with the
original material critically than students who take the summary as an
adequate representation.
Toward Accountable Summarisation
The path toward more accountable AI summarisation of sacred and
ideologically charged texts runs through transparency, consultation, and
epistemic humility. Transparency means AI systems that disclose their
interpretive assumptions and the characteristics of their training data when
producing summaries of culturally sensitive material. Consultation means
involving faith communities, religious scholars, and cultural experts in the
evaluation of AI outputs before they are deployed in contexts affecting those
communities. Epistemic humility means AI systems that acknowledge the limits
of their competence rather than producing confident-sounding summaries that
obscure how much they are missing.
A framework
for AI and theological diversity published in the Journal of
Religious Studies recommends multi-faith expert audits of AI systems before
deployment in academic or community settings, standardised disclosure
requirements for AI-generated religious content, and ongoing evaluation
processes that involve communities rather than just technical reviewers.
These recommendations apply equally to politically charged material: the
communities most affected by how AI handles their texts must have meaningful
input into how those systems are designed and evaluated.
Sacred texts survived millennia before AI. They will not be
destroyed by AI summarisation. But the communities whose lives are shaped by
those texts deserve AI systems that treat them with the complexity and
respect they warrant, not systems that reduce them to statistics and call the
output an understanding.
The connections here extend to how AI handles the linguistic
traditions in which many of these texts are embedded. As explored in The
Forgotten Accent, AI systems systematically underperform on
minority languages and non-dominant linguistic registers, compounding the
representation problem in sacred text processing. A model that struggles with
Quranic Arabic, classical Hebrew, Pali, or Sanskrit is not equipped to
summarise the texts written in those languages with the fidelity they
require. Addressing AI’s handling of sacred and ideologically charged texts
requires addressing its linguistic limitations alongside its interpretive
biases.
About the Author
Stuart Kerr is the Technology Correspondent for LiveAIWire,
covering artificial intelligence, ethics, and the ways technology is
reshaping everyday life.