E-Prosiding Persidangan Serantau Manasik Haji 2025 (MANASIK2025)

MANASIK2025 Putrajaya, Malaysia, 7-8 Oktober 2025 632 must bridge. Hajj and ʿUmrah as safety -critical, context-dependent use cases Hajj/ʿUmrah rulings are uniquely sensitive because mistakes directly compromise the validity of rites and may entail compensations ( dam , fidya ). Malaysia’s National Hajj Muzakarah (under Lembaga Tabung Haji, TH) has, for decades, produced practical resolutions (e.g., sequencing errors, health- related dispensations, travel contingencies) tailored to Malaysian pilgrims’ recurring challenges (Lembaga Tabung Haji, n.d.). This corpus complements general fiqh with operational guidance anchored in maslahah, risk mitigation, and local capacity constraints — guidance that TH operationalizes through courses and on-the-ground counselors. Regionally, the Saudi ecosystem has also digitized pilgrim services through the Nusuk platform, consolidating permits, logistics, and guidance into unified user flows (Saudi Ministry of Hajj and Umrah, n.d.). At the frontier, Saudi authorities have even tested a “fatwa robot” at the Two Holy Mosques to route users to remote muftis and curated content — an experiment that highlights both the promise and sensitivity of automating religious Q&A in sacred spaces (Saudi Press Agency, 2022) Strengths/limitations. The Hajj-specific literature and documentation provide rich, recurrent scenarios ideal for computational modeling (e.g., ritual states, time windows, contingencies). However, the materials are often scattered across PDFs and advisories with heterogeneous metadata. There is little formal evaluation of user understanding or error reduction due to decision aids — an evidentiary gap for claims about “safer” guidance at scale. Risks of generative AI in safety-critical advice Outside the religious domain, AI scholarship has converged on well-documented risks. Large language models (LLMs) can “hallucinate” facts, misstate sources, and fabricate citations, with error profiles that vary by task and prompt (Ji et al., 2022, 2024). Position papers caution that scale alone does not solve these issues and may amplify opaque biases and spurious fluency (Bender et al., 2021). Industry reporting similarly emphasizes that hallucinations remain structurally unavoidable, spurring interest in retrieval-augmented generation (RAG) and layered evaluation as partial mitigations. In governance research, algorithmic audit frameworks highlight the need for documentation, pre-deployment testing, and incident response for high-stakes uses (Raji et al., 2020). Strengths/limitations. This literature offers concrete design patterns — provenance-binding, constrained decoding, audits — that are directly applicable to fatwa automation. However, it is domain-generic. It does not specify how to encode jurisprudential hierarchies (e.g., sihhah of rituals, rukhsah conditions), nor how to reconcile multiple madhāhib and federal– state authority boundaries — precisely the issues that define Malaysian Hajj/ʿUmrah guidance. Retrieval-augmented generation and provenance as enabling techniques RAG architectures combine parametric generation with non-parametric retrieval from a curated corpus, improving factuality and enabling source attributions in the response (Lewis et al., 2020). For fatwa automation, a RAG stack could index (1) state-gazette d fatāwā ( sighah + explanatory notes), (2) national Muzakarah decisions, and (3) TH Hajj resolutions — tagged by jurisdiction, status, and ritual phase — so that generated answers always cite the exact underlying text and signal its authority level. This architecture aligns with institutional practice (start from authoritative text before issuing new ijtihād ), and it supports audit trails for quality review. Strengths/limitations. RAG is empirically promising and operationally intuitive for institutions. Yet retrieval quality depends entirely on curation. If state portals differ in metadata quality or publication