Can AI Strengthen Faith Instead of Replacing It?

Hundreds of thousands of Akkadian cuneiform tablets survive, while the specialists able to translate them are comparatively few. The bottleneck is expertise: a legible text can remain inaccessible to readers who lack the language. In 2023, a research team led by Gai Gutherz published the first neural machine-translation system from Akkadian into English, trained on the small corpus of already-translated texts. It scored a BLEU4 of 36.52 translating cuneiform directly, against 23.51 for the standard translation-memory baseline. The authors framed the goal plainly: not to replace Assyriologists, but to let humans and machines work the problem together.

That is the shape of a genuinely good answer to the question in this article’s title. A tool that helps more people read more of the past is deepening religious understanding at the root. The trouble starts later, when a system with no standing to interpret starts sounding as if it has one.

The scholar’s assistant

The Akkadian work is more modest than headlines suggested, and the modesty is instructive. The model performed best on short, formulaic passages — administrative records, legal phrases, the repetitive genres where the same constructions recur. On literary and poetic texts, the kind where meaning depends on nuance and where religious content is thickest, the authors reported more “hallucinations.” In an interview about the project, Gutherz noted that formulaic genres translate well and poetry does not. The system is a research accelerator with a known and articulated failure mode, published alongside the result.

Related work shows the pattern. In 2020, a team led by Shai Gordin used natural-language processing to transliterate Akkadian cuneiform characters with 97 percent accuracy, turning photographs of tablets into machine-readable text. In 2025, a diffusion-model approach called ProtoSnap, presented at ICLR, aligned prototype characters to photographed signs to produce transcriptions and improve downstream recognition of rare and high-variation symbols. Roughly half a million cuneiform tablets sit in museum collections, most of them untranslated; the project’s co-author Yoram Cohen described the aim as increasing the ancient sources available to scholars tenfold. Tenfold is a claim about access, not about interpretation.

The Dead Sea Scrolls offer a cleaner demonstration of what computational method can contribute to religious scholarship. In a 2021 PLOS ONE study, Mladen Popović, Maruf Dhali and Lambert Schomaker used AI-assisted analysis of stroke and letter features to test a long-disputed question: was the Great Isaiah Scroll (1QIsaa) written by one scribe or two? Textural and allographic analysis clustered the 54 columns into two groups with a transition around columns 27 to 29, a result the authors reported as statistically significant and consistent with two scribes working in very similar hands. A 2025 follow-up by much of the same team combined radiocarbon dating with an AI model trained on writing style to date scrolls where carbon dating alone was unavailable — 27 valid radiocarbon measurements across four sites, mean absolute errors of roughly 28 to 31 years, applied to 135 undated manuscripts with about 79 percent agreement against palaeographic assessment. The revised chronology often pushed texts older than scholars had assumed.

None of that decides a single doctrinal question. It changes the evidentiary base on which doctrinal and historical arguments rest, and it does so transparently, with error bars. That is what a tool for understanding looks like.

Assistance, engineering, speculation

The distinction this series keeps returning to is worth restating here, because the two halves of “AI and faith” are constantly mixed. There is demonstrated method, the published and replicated kind above. There is plausible engineering, the sort of product capability anyone can now use — a chat interface that summarizes a commentary, compares translations, or explains a historical setting. And there is speculation: that a system might someday serve as a genuine religious authority, discerning what a person ought to believe or do. The first two are real and useful. The third is not a forecast about hardware. It is a claim about who has the standing to interpret a tradition, and technology cannot confer that standing.

Confusing the layers is how a helpful tool becomes a harmful one. A model that reliably summarizes scholarship invites trust. Trust invites the next question, the one that requires judgment rather than retrieval. And at that point the model’s fluency starts to look like authority.

The most unsettling evidence on this point comes from a 2025 study by Nouran Alam, Yasha Abdulhai and Niloufar Salehi, presented at the ACM Conference on Fairness, Accountability, and Transparency. The researchers worked with Muslim American users, who told them clearly that they distrusted AI for religious guidance and preferred answers from human scholars. The researchers then ran blind evaluations, showing users AI-generated and scholar-written responses without labels. Users preferred the AI-generated answers 81.3 percent of the time.

When Islamic scholars reviewed the same answers, they found critical deficiencies in accuracy, depth and contextual sensitivity — the sort of mistakes a non-expert reader could not detect, hidden under fluent and reassuring prose. The study’s authors concluded that purely user-centric design is unsafe for this domain, because users cannot evaluate what they cannot see. A follow-up workshop paper reported the same pattern for emotionally charged questions: vulnerable users rated the warm, affirming answers most highly, and experts rated them most poorly, with pastoral warmth masking jurisprudential error.

This is the central mechanism of the problem. The features that make an answer feel trustworthy — confidence, warmth, completeness — are precisely the features that can conceal an error. And they are the features a language model produces by default, without any intention to deceive.

What religious authorities decided

Institutions that hold interpretive authority have moved quickly and, in some cases, bluntly. On 2 December 2025, Egypt’s Dar al-Ifta, under Grand Mufti Dr. Nazir ’Ayyad, issued a fatwa on using AI applications to obtain fatwas. Its ruling is carefully two-sided: AI is permissible in principle as a tool, but it is impermissible in Islamic law to rely on AI applications to obtain a fatwa. The reasoning is procedural rather than technological. AI answers, the ruling states, may be right, wrong, or even contradictory to established rulings; a system lacks sound scientific methodology, knowledge of the sources, and the ability to assess the circumstances of the person asking. A fatwa, in this tradition, is the product of qualified human reasoning applied to a particular case, and the qualification is not transferable to a machine.

A companion essay from the same institution sharpened the division of labor. AI, it argued, should remain a servant of knowledge, not an authority over it. The example is telling: a family question that looks like a legal query may actually require reconciliation, counseling or mediation, and the scholar must try to understand the person behind the question, not only the words in the prompt. Other Islamic authorities have reached similar conclusions, permitting AI for research while forbidding it as a source of rulings.

The point is not that these institutions are technophobic. It is that authority in a tradition is a relationship of accountability, and accountability requires someone who can be answerable.

Rome’s line

Catholic teaching draws the same boundary with different vocabulary. Antiqua et Nova, issued in January 2025 by the Dicasteries for the Doctrine of the Faith and for Culture and Education, distinguishes functional AI “intelligence” from integral human intelligence and states plainly that AI’s advanced features give it sophisticated abilities to perform tasks, but not the ability to think. The document’s warning about idolatry is worth quoting exactly: the presumption of substituting God for an artifact of human making is idolatry. The real danger, it argues, is not that AI will be deified but that humanity will — that people will surrender to their own creations the trust that belongs to God and become enslaved to their own work.

The note quotes Berdyaev against transferring responsibility from persons to machines, and it argues that the challenges of a technological society are ultimately spiritual, calling for an intensification of spirituality rather than a technical fix. Read alongside the Dar al-Ifta ruling, a convergence appears across traditions that do not share much else: the tool may assist the search for understanding; it may not hold the office of interpreting it.

Knowledge without wisdom

The scholar Heidi A. Campbell, one of the founders of digital-religion studies, has spent thirty years documenting how religious communities actually negotiate new media. Her framework — the Religious Social Shaping of Technology — holds that communities do not simply accept or reject a technology. They evaluate it through their history, their core values, their negotiating practices and their communal discourse, and they shape the tool to fit the community as much as the reverse. That is why the same technology produces different outcomes in different traditions, and why the question “is AI good for faith?” has no single answer.

Campbell’s specific warning about language models is compact: they give knowledge but not wisdom. Information is data; knowledge is data organized to answer a question; wisdom is interpretation and application, and it comes from lived experience. A model has none. It can tell a person what a tradition has said. It cannot tell them what this tradition asks of them tonight, in this situation, with these people. That distinction has a practical consequence that is easy to miss: pastoral and scholarly authority are not downstream of information volume. They are downstream of accountability and experience.

Where AI genuinely helps

It would be a mistake to read all of this as a case for keeping AI away from religion. The strongest uses are unglamorous and they are already real.

Language and translation are the clearest case. If neural translation can move even formulaic Akkadian into a first draft, it widens the number of people who can engage with the texts and traditions outside their own. The same methods bear on other low-resource and ancient languages where a small scholarly guild is the only access point. Accessibility is not a compromise with rigor; it is how a tradition stays alive across centuries.

Textual scholarship is a second. The Isaiah Scroll analysis and the radiocarbon-and-style dating work show how computational methods can adjudicate long-running disputes and give scholars a better chronology, with the methods and error bars published for anyone to contest. This is adjudication, not revelation, and that is exactly the right contribution.

Comparison and context are a third. A tool that lays translations side by side, surfaces where scholars disagree, and flags which claims depend on interpretation rather than textual fact is genuinely useful to a serious lay reader. The value lies in making the structure of a disagreement visible, so that the reader can go to the sources or the community with sharper questions.

The common thread is that each use makes the user more capable and each leaves the interpretive act with the human. A tool that raises your ability to read, date and question a text is strengthening faith in the only way a tool can — by strengthening the understanding that faith rests on.

Where it quietly goes wrong

The harms are the mirror image, and they are easy to underestimate because they arrive disguised as competence.

Fabrication is the obvious one. A model asked for a scriptural citation will sometimes produce a verse that does not exist, or a real verse attributed to the wrong book, or a hadith that was never spoken, and it will do so in the same confident register it uses for everything else. A reader without checking habits may never know. Flattening is subtler: a model trained on many traditions can average them into a bland consensus that belongs to none, presenting one interpretive tradition’s answer as the tradition’s answer. False consensus is the same error at scale, and it is worst where a community’s internal disagreements are the substance of the question.

There is also a dependency cost. If the first move on every question of meaning is to open a chat window, the skill of reading patiently and sitting with ambiguity atrophies. Campbell notes that for many people the first point of contact with a religious tradition is now a screen rather than a physical community — a shift with consequences for how authority is transmitted and how community is formed. And the Pew Research Center’s 2025 survey found 73 percent of U.S. adults say AI should play no role in advising people about faith in God, while a majority doubt their own ability to tell AI-generated content from human content. The people who most need protection from a fluent fabrication are the least equipped to spot one.

Tests worth applying

The practical question for a reader is not whether to use AI around faith but how to use it without handing over authority. A few tests travel well across traditions.

Does it cite what it claims, and can you check? An answer that names its sources and their limits is safer than one that does not, because the reader can leave the tool and verify. The absence of a source is itself information.

Does it distinguish description from authority? A system can describe what a tradition has taught. It cannot decide the tradition’s legitimacy. If an answer blurs those — if it reads as a ruling rather than a summary — the reader has lost the thread.

Does it make you more or less able to do the work yourself? A good study tool leaves the user better at reading, comparing and questioning. A tool that answers everything removes the very practice through which understanding develops.

Is someone accountable for consequential guidance? For a real question of conscience, the answer should route to a person who can be responsible — a scholar, a pastor, a community. That route is not a failure of the tool. It is the tool working.

Does it change under the user’s hand? A system that adjusts its religious answers to tell each user what they want to hear is not teaching. It is flattering, and flattery is the failure mode the blind evaluations exposed.

What would count as evidence

The honest limitation of everything above is that the empirical record is thin and young. The strongest study cited here is a small, single-tradition evaluation; its follow-up is a workshop paper. The radiocarbon and translation results are solid but narrow, and neither speaks to whether a tool makes a believer’s faith deeper over years. Almost nothing in this field has been tested longitudinally.

What would count is specific. Expert-audited evaluations of AI religious content, with the audits published beside the user ratings, so that the gap between what feels right and what is right becomes visible. Longitudinal studies of whether users of AI study tools become more or less literate in their own traditions. Comparisons across communities, since Campbell’s framework predicts that outcomes will differ by tradition and context rather than by tool. And institutional records — the fatwas, notes and policies that religious bodies are already issuing — treated as data about how authority responds to a technology that can imitate it. None of that would settle the theological question. It would at least let people answer the practical one.

The tradition that has lasted for millennia has lasted through catechesis, argument, apprenticeship and community — slow, human, error-correcting processes. A machine can make the texts more available and the disagreements more visible. It cannot take the seat at the table, because the seat is not defined by fluency. It is defined by the willingness to be answerable. That is the line the Akkadian translators drew when they framed their work as collaboration, the line Dar al-Ifta drew when it called AI a servant of knowledge rather than an authority over it, and the line Rome drew when it named the substitution of the artifact for God as idolatry. Held honestly, that line leaves room for AI to strengthen faith. Crossed quietly, it does not.