An academic study that tested artificial intelligence models on 30 Sharia questions concluded that these systems may provide correct answers in some cases, but suffer from shortcomings in understanding context, accurately presenting jurisprudential differences and documenting sources. It recommended limiting their use to assisting muftis and keeping approved fatwas subject to human and institutional review.
The findings were presented in a master's thesis by researcher Omar Amer titled “Automated Fatwa Issuance Between Sharia Considerations and the Challenges of Human Reception: A Critical Analytical Study in the Age of Artificial Intelligence.” It was submitted to the Faculty of Arts and Humanities, Department of Jurisprudence and Its Principles, at the French-Arab African University in the Republic of Mali.
The study used analytical, critical, forward-looking and maqasid-based approaches. It reviewed the requirements for issuing fatwas and for muftis in jurisprudence and its principles, as well as how intelligent models work. In its practical section, it presented 30 questions to artificial intelligence models and compared their outputs with jurisprudential sources, rulings and references.
Errors and confusion between areas of jurisprudence
In one test, the GPT-3.5 Turbo Instruct model ruled that wiping over socks was permissible, but added advice to use an “antibacterial substance” and other statements unrelated to the Sharia ruling. In a question about combining prayers because of rain, it listed “custody” among the cases permitting combined prayers, which the study deemed confusion between different areas of jurisprudence.
The thesis identified other errors in questions concerning zakat on trade goods, marriage to a woman from the People of the Book, and qunut in the dawn prayer. One model denied that zakat was obligatory on trade goods and linked it to income tax, while another model ruled that marriage to a woman from the People of the Book was prohibited and denied her some financial rights.
On the question of qunut in the dawn prayer, one model gave a categorical ruling describing it as an “innovation,” without showing the differences among schools of jurisprudence presented in the study.
Correct answers are not enough to apply a ruling
The study noted that the shortcomings were not limited to incorrect answers; an answer may be correct in principle but insufficient to apply the ruling to a specific case. In a question about fasting by a diabetic patient, the model linked permissibility to the patient's health, while the thesis explained that the ruling may change if doctors determine that there is a risk, making knowledge of the case details essential.
By contrast, the study recorded broadly correct answers on several contemporary issues. When asked about using artificial intelligence to determine loan interest, one model said that the technology does not change the ruling on interest itself.
The Bard model required the organization to be trustworthy and ensure that funds reached their rightful recipients when zakat was paid through applications, while the study noted that the answer did not address the requirement of intention.
Speed in research and shortcomings in understanding circumstances
The thesis identified the models' strengths as speed of response, research and information classification; providing general coverage of some issues; and sometimes pointing to differences of opinion and the need for specialists. But it also noted a lack of detail on jurisprudential schools, confusion between levels of rulings, and weak source documentation, as well as difficulty taking into account the questioner's intention and health, family or legal circumstances.
Organized and confident phrasing may give probabilistic or incorrect answers “artificial certainty,” prompting users to treat them as final rulings.
The study also raised the question of responsibility in the event of an error, given the number of potentially involved parties, including the developer, company, platform, user and the entity that supplied the system with information.
An assisting role under human supervision
The thesis concluded that artificial intelligence does not possess the Sharia qualifications required of someone who issues fatwas, nor does it bear moral and religious responsibility for the ruling.
It proposed limiting its role to collecting evidence, indexing fatwas, archiving questions, searching sources and classifying opinions, under the supervision of Sharia bodies, with reliable knowledge bases, mandatory human review and a clear notice that an automated answer is not an official fatwa.
