QCRI/islamic-knowledge-ai-survey

Advances in AI Systems on Islamic Knowledge Capabilities: A Critical Survey

   

A comprehensive systematic survey of 160+ papers (2016–2026) examining how AI systems operationalize Islamic knowledge, spanning NLP, information retrieval, speech processing, multimodal learning, educational technology, and LLM alignment.

<p align="center"> <img src="https://gagan3012.github.io/islamic-knowledge-survey/static/images/fig2temporalnew%20(1)-1.png" alt="Publication trends" width="700"> </p>

Abstract

AI systems are increasingly mediating how Islamic communities access, study, and apply Islamic sources; still, research on Islamic-knowledge capabilities remains fragmented across NLP, information retrieval, speech, multimodal learning, educational technology, and recent LLM alignment work.

This survey presents a critical systematic review of 160+ papers from the past decade that incorporate Islamic knowledge in Machine Learning/AI. We propose a layered taxonomy that separates an epistemic view of Islamic knowledge (authority-bearing foundations and established disciplines) from an instrumental AI task layer (data and corpora, retrieval and grounding, understanding, reasoning support, evaluation and governance, and multimodal methods), while treating normative concerns as cross-cutting constraints.

Using this framework, we synthesize trends in datasets, benchmarks, and system architectures, highlighting the shift toward retrieval-grounded LLM pipelines, verification and deferral mechanisms, and emerging multimodal recitation and manuscript-processing systems. We also consolidate evaluation practices for trustworthiness, including provenance and faithfulness, disagreement-aware and school-of-thought-sensitive framing, calibrated abstention under underspecified queries, and safety and bias assessment for Islamic contexts.

Key Contributions

• —Layered Taxonomy — A two-layer framework separating the epistemic view of Islamic knowledge (Qur'an, Hadith, Fiqh, Theology, etc.) from the instrumental AI task layer (retrieval, grounding, reasoning, evaluation, multimodal methods).
• —Systematic Review (PRISMA-ScR) — Rigorous screening of 1,743 initial records down to 160 included studies, following the PRISMA-ScR framework for transparency and reproducibility.
• —Cross-Cutting Normative Dimensions — Analysis of doctrinal integrity, disagreement-aware framing, and deployment safety as cross-cutting concerns.
• —Comprehensive Papers Database — A searchable, filterable collection of all surveyed papers with metadata on domains, tasks, and research areas.

Taxonomy

<p align="center"> <img src="https://gagan3012.github.io/islamic-knowledge-survey/static/images/figsunburstv3.png-1.png" alt="Sunburst taxonomy visualization" width="600"> </p>

Epistemic Layer

AI Task Layer

Research Questions

1. 1.RQ1 — Domains & Tasks: What Islamic knowledge domains and application tasks have been operationalized in ML/AI systems, and how is this work distributed across subfields?
2. 2.RQ2 — Resources & Measurement: What datasets, benchmarks, and knowledge resources are available, and what assumptions do they encode about evidence, provenance, and interpretive diversity?
3. 3.RQ3 — Evaluation & Trustworthiness: How do studies evaluate trustworthiness, especially source faithfulness, doctrinal correctness, pluralism-aware answering, and safety/bias?

Methodology

We followed the PRISMA-ScR (Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews) framework:

<p align="center"> <img src="https://gagan3012.github.io/islamic-knowledge-survey/static/images/qcri%20project-prisma.drawio-1.png" alt="PRISMA flow diagram" width="600"> </p>

• —Sources: Semantic Scholar, IEEE Xplore, ACM Digital Library, ACL Anthology, arXiv
• —Coverage: 2016–2026
• —Screening: 1,743 initial records → 160 included papers

Key Findings & Challenges

Engineering Priorities

• —Provenance-preserving grounding — Retrieval-grounded pipelines with verifiable citations
• —Disagreement-aware systems — Present alternative scholarly views with supporting evidence
• —Calibrated abstention — Defer to qualified authority when grounding is unreliable
• —Interdisciplinary collaboration — AI researchers, Islamic scholars ('ulama), and community stakeholders
• —Benchmark investment — Evaluation protocols that penalize fabricated citations, with disagreement-aware scoring
• —Safety-first deployment — Islamic-specific red-teaming, bias checks, and governance frameworks

Citation

If you use this survey in your research, please cite:

@article{Bhatia_2026,
  title     = {Advances in AI Systems on Islamic Knowledge Capabilities: A Critical Survey},
  url       = {http://dx.doi.org/10.36227/techrxiv.177155997.77147487/v1},
  DOI       = {10.36227/techrxiv.177155997.77147487/v1},
  author    = {Bhatia, Gagan and Mubarak, Hamdy and Hawasly, Majd and Jarrar, Mustafa and
               Mikros, George and Zaraket, Fadi and Alhirthani, Mahmoud and Al-Khatib, Mutaz and
               Cochrane, Logan and Darwish, Kareem and Yahiaoui, Rashid and Alam, Firoj},
  year      = {2026},
  month     = feb
}