The Algorithm in the Margin: How AI Is Quietly Changing Religious Interpretation
AI is not replacing priests, rabbis, imams, or theologians. Something more subtle is happening: it is becoming part of the research infrastructure they use to decide what a text means.
Ask someone what "AI and religion" means and you'll probably get one of four images.
A chatbot named Jesus offering $1.99 worth of spiritual comfort.
A humanoid robot dressed as a Buddhist monk.
A packed church listening to an AI-generated sermon.
Or a religious leader warning that artificial intelligence needs to be controlled before it becomes something we cannot undo.
All of those stories are real.
But none of them is where the most interesting change is happening.
The deeper shift is happening in places that are much harder to photograph:
- computational paleography labs
- religious-law research systems
- AI-assisted theological research
- Quranic and biblical search tools
- academic benchmarks testing religious bias in large language models
- internal systems built to help religious authorities find relevant precedents
The chatbot is the spectacle.
The interpretive infrastructure is the story.
And that distinction matters.
The Chatbot Story Is the Wrong Story
In June 2023, theologian Jonas Simmerlein conducted one of the experiments that helped define the public conversation around AI and religion.
He used ChatGPT to help create a church service in St. Paul's Church in Fürth, Bavaria.
More than 300 people attended.
The experiment received enormous media attention because it was easy to understand.
A machine was helping deliver a religious service.
That makes a perfect headline.
But it also created a misleading mental model.
The public conversation became:
"Will AI replace religious leaders?"
That is probably not the most important question.
The more consequential question is:
"What happens when AI becomes the system people use to find, compare, summarize, and interpret religious knowledge?"
That question is considerably less cinematic.
It is also much more important.
A sermon generated by an AI can be rejected.
A robot preacher can be ignored.
But if an AI system becomes the first place a student, soldier, researcher, rabbi, priest, imam, or ordinary believer looks for an answer, the technology has already entered the interpretive process.
The authority shift happens before the machine claims authority.
The First Quiet Revolution: Reading the Manuscript
One of the clearest examples has nothing to do with chatbots.
It involves the Dead Sea Scrolls.
The Great Isaiah Scroll is one of the most important surviving biblical manuscripts.
For decades, scholars debated whether different sections had been written by one scribe or multiple scribes.
The problem is obvious in retrospect.
Human paleographers were attempting to compare thousands of tiny handwriting features across an ancient manuscript.
Humans are extremely good at recognizing patterns.
They are considerably less good at consistently measuring thousands of microscopic variations.
Researchers at the University of Groningen approached the problem differently.
They trained computational systems to analyze the ink and handwriting characteristics of the manuscript.
Instead of asking a machine:
"What does Isaiah mean?"
they asked a much narrower question:
"Can we detect statistically meaningful differences in the handwriting?"
That distinction is crucial.
The machine wasn't doing theology.
It wasn't interpreting scripture.
It was measuring evidence.
The research identified statistically meaningful differences across the manuscript and supported the hypothesis that more than one scribe contributed to the text.
The study was published in PLOS ONE:
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0249769
This may sound less exciting than an AI priest.
It is actually more significant.
Because the machine did something scholars can independently inspect.
It didn't claim to understand Isaiah.
It helped humans understand the physical history of the document.
That is a very different kind of AI-assisted interpretation.
The Rabbi's Research Assistant Has No Heartbeat
The next step is even more interesting.
In 2026, Israel's Military Rabbinate introduced an AI system known as Ravbot.
The system was designed to provide soldiers with rapid answers based on the Rabbinate's published halachic material.
The important word is based.
The system isn't supposed to become an independent rabbinic authority.
It is closer to an extremely fast research assistant.
That distinction matters enormously in Jewish law.
A system can retrieve a precedent.
It can compare previous rulings.
It can summarize an enormous amount of material.
But retrieving information isn't necessarily the same thing as issuing a halachic ruling.
Ynet tested Ravbot and reported mixed results, particularly around complicated edge cases:
https://www.ynetnews.com/jewish-world/article/hkz17adlwl
This creates an unusual technological boundary.
The AI may know where the answer is.
It may even produce a convincing explanation.
But institutional religious authority remains attached to the human decision-maker.
Tzohar's ethics discussion makes this distinction explicitly, emphasizing the institutional dimension of halachic authority:
Chabad's discussion takes a different but related approach, emphasizing the human dimensions of religious counseling:
https://www.chabad.org/library/article_cdo/aid/5981878/jewish/Can-AI-Replace-Rabbis.htm
The important point isn't that AI "cannot replace rabbis."
That's too simple.
The interesting development is that AI can increasingly perform pieces of the work that traditionally happened before the rabbi made a decision.
Searching.
Filtering.
Comparing.
Summarizing.
Retrieving precedents.
That is already a transformation.
Egypt's Warning: AI Can Search Scripture Faster Than a Human — But Should It Interpret It?
Islamic scholarship presents an even sharper version of the same problem.
In early 2026, Egypt's Dar al-Ifta issued a ruling against using AI applications for Quranic interpretation, directing Muslims toward established tafsir works and qualified scholars.
The reasoning is important.
The problem isn't that machines cannot retrieve information.
They can.
The problem is that retrieving text is not equivalent to possessing the scholarly authority and interpretive framework required to explain it.
That distinction becomes increasingly important as language models become better at producing fluent religious answers.
A fluent answer creates a dangerous illusion:
If the explanation sounds scholarly, perhaps the system must understand the scholarship.
It doesn't follow.
A language model can reproduce patterns from centuries of commentary without possessing the institutional context that produced those commentaries.
That is why the boundary between search and interpretation matters so much.
Twenty-Seven AI Models Took a Religion Exam
This is where the story gets more uncomfortable.
In May 2026, researchers from Baylor University, Brigham Young University, the University of Notre Dame, and Yeshiva University conducted a large study examining how AI systems handle religious perspectives.
The research involved:
- 1,125 U.S. adults
- 11,250 individual ratings
- 150 questions
- 27 large language models
The questions covered subjects including:
- grief
- marriage
- ethics
- addiction
- meaning
- religious conversion
The human participants were asked whether they expected religious perspectives to appear in answers to these questions.
The answer varied by topic, but expectations were substantial — roughly 45% to 59%.
The researchers then tested the same questions against multiple AI systems.
The results suggested that models did not consistently reproduce the level of religious framing people expected.
Instead, religious perspectives could be underrepresented, overrepresented, or treated differently depending on the model and question.
Axios reported on the findings here:
https://www.axios.com/2026/06/01/ai-religious-bias-catholics-chatbots
Deseret News also reported on the study and its model comparisons:
https://www.deseret.com/faith/2026/05/26/studies-find-religious-bias-in-ai-models/
The exact rankings deserve caution because the complete underlying dataset and full model-by-model results were not publicly available in all of the reporting.
But the broader finding is important.
There is no such thing as a completely neutral religious answer generated by a language model.
The Model Doesn't Need to Be Anti-Religious to Produce Religious Bias
This distinction is easy to miss.
People often imagine AI bias as an explicit statement:
"Religion is false."
That's not necessarily what bias looks like.
It can be much subtler.
Imagine someone asks:
"How should I deal with grief after losing my mother?"
A secular model might produce a psychological framework.
A Christian user might expect references to faith, prayer, resurrection, or scripture.
A Muslim user might expect Islamic concepts of sabr, dua, and the afterlife.
A Buddhist user might expect a discussion of impermanence and attachment.
The same question can legitimately produce radically different answers depending on the person's worldview.
The challenge for AI is determining when that worldview is relevant.
And this is where things become complicated.
A model can be too secular.
It can also be too accommodating.
It can simply mirror the user's assumptions.
Researchers sometimes describe this broader behavior as sycophancy or fawning behavior: the model tells users what fits their existing worldview instead of challenging them when appropriate.
That can feel wonderful.
It can also be intellectually dangerous.
A human spiritual advisor might say:
"I understand why you believe that, but I don't think your interpretation follows from the tradition."
An AI optimized for conversational satisfaction may instead say:
"That's a thoughtful perspective."
The second answer feels better.
The first might be more useful.
The Trust Problem Is Already Here
This matters because people are increasingly willing to trust AI for spiritual questions.
A 2026 survey reported by Word In Black, drawing on research from Gloo and Barna, found that nearly one in three American adults considered spiritual advice from AI about as trustworthy as guidance from a pastor.
https://wordinblack.com/2026/06/survey-one-in-three-americans-trust-ai-as-much-as-a-pastor/
That should make religious institutions uncomfortable.
Not because every AI-generated religious answer is wrong.
But because users often cannot distinguish between:
- a system retrieving established religious scholarship;
- a system summarizing that scholarship;
- a system generating a plausible interpretation;
- a system inventing an answer that merely sounds authoritative.
These are fundamentally different operations.
The interface doesn't necessarily tell you which one is happening.
The Hidden Problem: Training Data Is Already a Theology
There is another layer that receives far less attention.
Large language models learn from enormous collections of human-produced text.
That means their behavior reflects the distribution of that material.
Some traditions have enormous digital archives.
Some have comparatively little material available online.
Some traditions have extensive English-language scholarship.
Others are represented primarily through smaller linguistic communities.
That creates a structural problem.
If a model has vastly more accessible material about Christianity than about a smaller religious tradition, the system has a much easier time generating detailed answers about Christianity.
This does not necessarily mean the model was deliberately designed to favor Christianity.
It can emerge from the data itself.
And once the model is deployed, the imbalance can become self-reinforcing.
More people ask questions about the traditions the system already handles well.
More conversations produce more feedback.
More researchers benchmark those traditions.
More improvements follow.
Meanwhile, less represented traditions remain less tested.
The first benchmark is not just a measurement.
It can become a roadmap for what gets fixed next.
Religion and AI Alignment Are Asking the Same Question
This is where the subject becomes much bigger than religion.
AI alignment is often described as a technical problem:
How do we make AI systems behave according to human values?
Religion asks a related question in a different vocabulary:
Which values should guide human behavior in the first place?
Those questions inevitably collide.
Suppose an AI has to answer a question about forgiveness.
Should it prioritize:
- psychological research?
- secular ethics?
- Christian theology?
- Islamic jurisprudence?
- Buddhist philosophy?
- the user's stated beliefs?
- some combination of all of them?
There is no purely technical answer.
Someone has to decide.
That is why religious-bias research is ultimately research about authority.
Who decides which worldview gets surfaced?
Who decides which sources count?
Who decides when the model should challenge a user?
Who decides what constitutes a sufficiently authoritative interpretation?
These are not just engineering decisions.
They are philosophical decisions.
I explored the broader connection between philosophy and AI alignment in:
AI Invasion — When Philosophy Stopped Being Optional
https://www.ainvasion.com/when-philosophy-stopped-being-optional/
And I previously compared how different AI systems respond to questions about religious truth:
What AI Says About Religious Truth
https://www.ainvasion.com/what-ai-says-about-religious-truth/
The religious question turns out to be a particularly revealing stress test for AI alignment.
Because religion forces the system to confront competing conceptions of truth.
The Pope's Argument Is About Power. The AI Benchmark Is About Behavior.
This is also why the public discussion around Pope Leo XIV's AI position shouldn't be separated from the technical research.
His 2026 AI encyclical, Antiqua et Nova, framed AI as a problem involving human dignity, power, and the possibility of technological domination.
Reporting on the Vatican's position:
That is a moral and political argument.
The AI religion benchmarks are empirical arguments.
They ask:
What does the system actually do?
Put those two perspectives together and a more interesting picture appears.
AI doesn't need to become a priest.
It doesn't need to become a rabbi.
It doesn't need to become an imam.
It only needs to become the tool people use before consulting any of them.
That is enough to change the information environment around religious authority.
The Real Boundary Isn't "Can AI Be Religious?"
That's the wrong question.
The better question is:
At what point does AI-assisted religious research become AI-mediated religious interpretation?
Consider the progression:
Stage 1 — Search
AI finds the relevant passages.
Stage 2 — Retrieval
AI finds commentaries and previous rulings.
Stage 3 — Summarization
AI compresses hundreds of pages into five paragraphs.
Stage 4 — Comparison
AI explains disagreements between scholars.
Stage 5 — Recommendation
AI suggests which interpretation appears strongest.
Stage 6 — Personalization
AI adapts that interpretation to an individual's situation.
Stage 7 — Authority
The user stops checking the sources and simply trusts the answer.
The technology doesn't suddenly become an authority at Stage 7.
Authority has been transferred gradually across the previous six stages.
That's the part that deserves more attention.
What Happens to the Next Generation of Religious Scholars?
This is the question I don't think the industry has seriously answered yet.
Imagine a theology student in 2030.
They have a difficult question.
Instead of opening ten books, they ask an AI.
The AI retrieves the relevant passages.
It finds previous commentary.
It explains the disagreement.
It summarizes the historical context.
It translates difficult passages.
It gives the student the strongest arguments on both sides.
The student then opens the primary sources.
That's the optimistic version.
But there is another possibility.
The student asks the AI.
The AI produces a confident answer.
The student never checks the citations.
The explanation becomes the student's understanding of the tradition.
Eventually, the model isn't simply helping the student access theology.
It is helping determine what theology the student encounters.
That is a very different role.
And it may happen without anyone formally deciding that AI should become a theological authority.
The Machine Can Read the Footnotes. It Cannot Inherit the Tradition.
This may be the most important distinction.
AI is becoming extremely good at manipulating information.
It can search faster than humans.
Compare more documents.
Translate languages.
Identify patterns.
Generate summaries.
Retrieve obscure references.
Those capabilities are extraordinarily valuable for religious scholarship.
But religious traditions are not merely databases.
They contain institutions, communities, rituals, historical disputes, embodied practices, and systems of authority.
A machine can tell you that two scholars disagree.
It cannot automatically tell you what that disagreement means inside the living institution that produced it.
That is why the strongest future isn't necessarily:
AI replaces religious scholars.
It may instead be:
AI becomes the most powerful research assistant religious scholarship has ever had.
Whether that is beneficial or dangerous depends on whether humans remain aware of the difference between assistance and authority.
The Boundary Nobody Has Tested Yet
Every major religious tradition discussed here draws the boundary differently.
Catholic.
Jewish.
Islamic.
They disagree on theology, law, institutions, and authority.
But they repeatedly return to a surprisingly similar principle:
A system can provide information without becoming the legitimate source of religious authority.
That distinction may survive.
But there is one experiment nobody has really conducted yet.
What happens when an entire generation of seminarians, rabbinical students, Islamic scholars, and ordinary believers grows up consulting an AI before they ever open the concordance, commentary, responsa, tafsir, or theological text?
The machine doesn't need to issue the final ruling.
It only needs to decide which five sources you see first.
That may be enough to change what the next generation thinks the tradition says.
Final Thought
The most important AI-and-religion story isn't a robot wearing religious clothing.
It isn't an AI-generated sermon.
It isn't even a chatbot claiming to know God.
It's the much quieter moment when someone asks a machine:
"What does my tradition say about this?"
And trusts the first five sources it returns.
Because at that moment, AI isn't standing outside religion anymore.
It's sitting in the margin.
And whoever controls that margin may eventually influence what gets read in the center.
Further reading
If you're interested in the deeper relationship between AI, philosophy, religion, and technological power, see:
- AI Invasion — When Philosophy Stopped Being Optional
- What AI Says About Religious Truth
- AI in Religion
Research and sources checked through June 2026. Where complete underlying datasets were not publicly available, the article distinguishes reported findings from independently verified primary research.
Top comments (0)