Grounded AI for Islamic questions: how 'no source, no answer' works and where it still fails

Why chatbots fail on religious questions

A general-purpose language model learns from an enormous mix of text: scholarly works, forum posts, marketing, translations of varying quality, and arguments between all of them. When you ask it a fiqh question, it produces the most plausible continuation given all of that. Three documented consequences:

1. Blended schools. The answer may stitch Hanafi, Shafi'i and Ja'fari positions into one confident paragraph that no school actually holds.
2. Permissive drift. A University of Malaya comparative study found a systematic bias toward permissiveness, with reliance on secondary sources rather than authoritative collections.
3. Invented citations. Asked for a source, a chatbot can produce a realistic-looking reference that does not exist.

None of these is a bug that a better model necessarily fixes, because the model is doing what it was built to do: continue the text plausibly. The fix is architectural.

How grounding changes the answer

1. A library is defined. Named sources only: for example, the English fatwa archive of Dar al-Ifta al-Misriyyah, the rulings published by the office of Ayatollah Sistani, named standards such as AAOIFI's. Each document carries its institution, school, date and link.
2. The question is used to retrieve. The system searches the library for passages relevant to the question, usually by meaning rather than exact words.
3. The school is established first. A well-built system asks which school or marja' you follow before answering, because the right passage depends on it.
4. The answer is generated from the retrieved passages only. The model is instructed to explain what the passages say and nothing more.
5. The passages are shown. Quote, source, date, link. You can read the original.
6. Nothing found means no answer. The system says it found no relevant source from your school and suggests asking a scholar.

The result is a reference tool with a conversational interface. Every answer passes the five-question check by construction, because the source, the school and the date come from the retrieved document and the decline behaviour is built in.

What grounding does not do

Grounding makes an answer checkable. It does not make it a ruling. Dar al-Ifta's fatwa 22255 bars relying on AI applications to obtain a fatwa; the reasoning concerns the nature of a fatwa as the act of an accountable, qualified human, and a better retrieval system does not alter that. A grounded tool's correct self-description is "here is what this scholar said about a question like yours", followed by "ask a scholar of your school about yours."

Known failure modes, honestly

A tool that tells you about these failures is more trustworthy than one that claims to have none.

Worked example: HalalScreener's Scholar

We run this site, and we run a grounded assistant, so it is the honest example to use. Here is how it is built and how to catch it out.

• Library. Fatwas from the English archive of Dar al-Ifta al-Misriyyah and rulings from the office of Ayatollah Sistani, among others, each tagged with institution, school and date.
• School first. If your profile does not record a school, Scholar asks before answering.
• No source, no answer. If retrieval returns nothing relevant from your school, Scholar says so and points you to a scholar. It is designed not to fill the gap from the model's memory.
• Sources shown. Each answer carries the passage and the link.
• Not a mufti. Scholar's answers are framed as references to named rulings, with the recommendation to consult a qualified scholar for anything that matters.

How to test it: ask something your school has not addressed in those archives and see whether it declines. Ask the same question twice in different words and compare. Open every link. If it fails any of the five questions, tell us at hello@halalscreener.app and we will fix it or document the limitation here.

Scholar is a HalalScreener product. HalalScreener screens stocks, ETFs and crypto; it does not certify, and it does not issue rulings.

The standard this implies

Grounding is one of the seven commitments in the Shariah Agent Standard, alongside school declaration, dating, decline behaviour, human hand-off, logging and a named accountable person. Any builder can adopt it. Any user can check it.