Using AI responsibly in Islamic apps: sources, scope and scholar review

Where AI appears in Islamic apps, and why it needs rules

AI now appears across Islamic apps, and each use carries a different level of risk. A Quran question-and-answer assistant, a hadith search tool, an Islamic studies tutor and a so-called fatwa bot may all use the same kind of model, but they should not be allowed to do the same things.

• Quran Q&A. Explaining a verse or a word. Lower risk when tied to a named tafsir; higher when the model answers from memory.
• Hadith search. Finding narrations on a topic. Useful, but dangerous if the model writes the narration text itself or states its grading.
• Learning assistants. Explaining tajweed rules, Arabic grammar or a lesson. Generally safe when grounded in a curriculum.
• "Fatwa bots". Answering personal religious questions about worship, marriage, money or family. This is where AI causes the most harm, because rulings depend on circumstances and qualified judgement.

The common failure is the same in all four: a fluent model fills gaps with confident text. In a shopping app that produces a wrong product description; in an Islamic app it can misquote revelation or misattribute a ruling. Guardrails are the product decisions that prevent that, and they start before any prompt is written. For the tafsir case in detail, see our guide to AI tafsir apps.

How to write a scope policy for an Islamic AI feature

A scope policy is a short document that sorts every expected question into one of three responses: answer, answer only with a named source, or decline and refer. It is written with the scholars advising the project, approved before development, and turned into the model's instructions and the test set.

The policy should also define the decline message itself. A good decline is short, respectful and useful: it explains that the app does not issue rulings, suggests asking a qualified scholar or a recognised fatwa body, and, where possible, points to the related verse or lesson the app can show.

Grounding answers in named sources, and never generating sacred text

The single most important rule is that a model never writes Quran or hadith text. Language models reproduce familiar wording well but not perfectly, and a missing word or a changed letter in a verse or narration is unacceptable. The safe architecture separates references from text.

1. Store sources as structured data. The Quran by surah and ayah; hadith by collection, book and number; tafsir passages by ayah range; each from a verified copy you have the right to use.
2. Retrieve before answering. The app finds the relevant passages and gives them to the model with the question.
3. Have the model cite by reference. The model's answer refers to items by their identifiers, such as a surah and ayah number or a hadith collection and number.
4. Resolve references in the app. The app checks that each reference exists and inserts the exact text from the database. An unknown reference is dropped, not guessed.
5. Show the source next to the answer. The reader sees the book, collection or tafsir name and can open the full original.

The same principle applies to hadith grading. Scholars grade narrations through careful study of chains and texts; an app should show the grading stated by a named scholar or source, if it stores one, and should never generate a grading of its own.

Handling scholarly differences and refusing rulings

Where qualified scholars differ, an Islamic AI feature should report the difference as its sources state it and should not choose a winner. Many questions of practice have more than one recognised position within mainstream scholarship, and a model that confidently presents one view as the only view misleads the user.

• Attribute every position. Say which book or scholar holds a view; never write "scholars say" without a source.
• Do not merge views. A summary that blends two positions into a third creates a view no one holds.
• Stay neutral in tone. Avoid language that ranks or dismisses positions, and avoid sectarian framing.
• Point to local guidance. For practice, users should follow the scholars and fatwa bodies they trust.

Refusing rulings is not a weakness of the product; it is part of its design. A fatwa depends on the questioner's exact situation, which an app cannot verify, and on the judgement of a qualified mufti. An Islamic AI feature should recognise these questions reliably, including when they are phrased indirectly, and decline every time, not only when it is phrased as an explicit request for a ruling.

Setting up scholar review and evaluation sets

Scholar review only works when it has a defined role, a defined workload and the authority to block a release. Before development, agree who the reviewers are, what they approve and what happens when they object.

• A small review board. Two or three qualified reviewers, ideally covering the disciplines the app touches, such as tafsir, hadith and tajweed.
• A written mandate. The board approves the scope policy, the source list and the decline wording, and signs off each release that changes AI behaviour.
• An evaluation set by category. Questions for each row of the scope policy, with the expected policy outcome and, where relevant, the expected source.
• Adversarial cases. Requests to quote a verse from memory, to grade a hadith, to issue a ruling, or to take sides between schools.
• A release gate. Any change to the model, instructions or sources re-runs the set, and failures in decline categories block the release.

In Aalim, the Quran learning app FLUIDS designed and built, FLUIDS's role included the AI evaluation pipeline alongside product design and the mobile build. Treating evaluation as a permanent part of the product, rather than a one-off check before launch, is what keeps behaviour stable when the underlying model changes.

Reporting wrong answers, logging and user privacy

Every AI answer in an Islamic app should carry a visible way to report it, and every report should reach a person who can act on it. Wrong answers will happen; what matters is that they are caught, reviewed by a qualified reviewer and fixed in the sources, instructions or evaluation set.

Logging needs a balance. The team needs enough to investigate a report: the question, the sources retrieved, the answer and the model version. It does not need everything forever. Religious questions can reveal personal circumstances and beliefs, so keep logs minimal, restrict access, set a retention period and never use conversations to train or tune a model without the user's explicit consent.

Requiring sign-in for AI features, as Aalim does for its AI tafsir conversations, can give users a history they control and give the team a way to handle reports and apply usage limits. In Saudi Arabia, personal data processing falls under the Personal Data Protection Law (PDPL), and national AI ethics guidance is published by the Saudi Data and AI Authority (SDAIA). Confirm the obligations that apply to your product with qualified counsel.

Transparency labels and a worked example

Users should always know what they are reading: revealed text, a named scholar's explanation or an AI-written summary. Three labels cover most cases, and they should be visible in the interface, not buried in terms of use.

• Source label. The Quran text, hadith collection or tafsir book the content comes from.
• AI label. A clear mark on any text written by AI, such as a simplified explanation.
• Scope statement. A short line explaining that the feature is a study aid, can make mistakes and does not issue rulings.

Aalim shows parts of this approach in a real product. Its tafsir uses named modes, labelled Al-Muyassar, Al-Mukhtasar and Al-Saadi, rather than an unnamed chat. Reading and AI assistance are separate experiences, so the Quran reader works on its own while AI tafsir conversations require a signed-in session and a connection. And the product states its purpose plainly: AI that guides, corrects and explains, while the traditional teacher stays where they are.

FLUIDS designs and builds Quran and Islamic learning apps with AI features for organizations and founders in Saudi Arabia and the GCC. See the Aalim case study and how we approach AI features inside a product. Our checklist of Quran app features covers the reader that should sit underneath any AI layer.