Conference Proceedings · Volume 1

From Human Labor to Machine Logic: Large Language Models and the Epistemic Structure of Ijtihād

Mohammad Abul Mufazzal
Faculty of Social Sciences, Ala-Too International University, Bishkej Kyrgyzstan
Correspondence: abulmufazzal.mohammad@alatoo.edu.kg

Naseem Fatima Chaudhary
Centre for Arabic & African Studies, Jawaharlal Nehru University, New Delhi, India

DOI: 10.68315/svpc26-u2q2vhc4
Publication date: November 20, 2026

This paper was presented at the Silicon Valley Peace Conference 2026 in Palo Alto, California, United States, held September 17–20, 2026, under the theme “AI’s Role in Community Development and Peacebuilding in Developing Nations,” and is published as part of the Proceedings of the Silicon Valley Peace Conference 2026.

Abstract

Large language models (LLMs) now produce Islamic legal guidance that is fluent, personalized, and, on its surface, hard to distinguish from the reasoned opinion of a trained jurist. This article asks what such systems do to the epistemic structure of ijtihād—the disciplined interpretive effort through which qualified scholars derive rulings from revealed sources—and argues that the widely voiced worry about “algorithmic religious authority” has been framed too loosely to do real work. Drawing on classical uṣūl al-fiqh and recent empirical evaluations of Islamic chatbots, the study makes three moves. First, it separates epistemic authority (being a reliable source of true legal propositions) from deontic authority (having the standing to oblige a questioner) and shows that, while LLMs may in principle narrow the epistemic gap, deontic authority is constituted by accountability (taklīf), certification (ijāzah), and communal recognition that no unaccountable system can bear. Second, it revisits the “probabilistic versus normative” contrast that dominates the field: because ijtihād classically yields ẓann (disciplined probable opinion) rather than certainty, the salient difference is not probability as such but the object over which the probability ranges—linguistic surface form for the model, the divine ruling for the jurist—together with the method and ownership attached to each. Third, it identifies two opposite failure modes that deployment risks producing: the collapse of legitimate disagreement (ikhtilāf) into an algorithmic monoculture, and, paradoxically, an anarchic talfīq in which lay users assemble self-serving rulings from decontextualized fragments. Reading the isnād system as a provenance protocol, the article explains why better retrieval solves the problem of transmission while leaving the problem of interpretation untouched. It closes with a design and governance framework—structured pluralism under human authority—that positions retrieval-augmented systems as research infrastructure for scholars while reserving the binding act of iftāʾ for accountable human agents.

Keywords: ijtihād, algorithmic authority, large language models, uṣūl al-fiqh, ẓann, ikhtilāf, retrieval-augmented generation, digital religion, Islamic ethics

Every major shift in Muslim communication technology has quietly rearranged the map of religious authority. The printing press put the Qurʾān and the legal commentaries into hands that had never touched a manuscript; the cassette sermon manufactured a new kind of preacher whose reach owed nothing to the seminary; satellite television produced celebrity muftīs; and the early internet scattered fatwās across websites that answered to no council (Bunt, 2009; Hirschkind, 2006; Messick, 1993). Each wave widened access and, in the same motion, loosened the grip of the institutions that had traditionally certified who was entitled to speak. Large language models (LLMs) are the newest wave, but they differ from their predecessors in a way that matters. Cassette tapes and fatwā archives circulated opinions that human scholars had already formed. A conversational model generates something new for each question—an answer shaped to the phrasing, situation, and apparent expectations of the person asking (Ahmad, 2025; Latifi, 2024). For the first time, a technology does not merely transmit the products of interpretive labor; it performs a convincing imitation of the labor itself.

This imitation is the source of both the promise and the problem. Islamic chatbots such as Ansari Chat, SheikhGPT, and a growing roster of state-backed, commercial, and open-source systems now offer believers instant, patient, plausibly sourced guidance in natural language (Ahmad, 2025). The convenience is real, and so is the unease it has provoked among scholars and fatwā bodies, who increasingly describe a phenomenon that this article, following Ahmad (2025), calls algorithmic religious authority: the perceived legitimacy that users grant to a computational system that produces normative religious guidance without answering to any of the classical mechanisms—isnād (the authenticated chain of transmission), uṣūl al-fiqh (the principles of jurisprudence), or ijmāʿ (scholarly consensus)—by which such authority has historically been earned and checked.

The trouble with the existing debate is not that this worry is misplaced but that it is stated too bluntly to guide either scholarship or design. Most treatments pose the question as a stark opposition—human, normative, and accountable reasoning on one side; machine, probabilistic, and unaccountable calculation on the other—and then conclude, unsurprisingly, that a machine that fabricates facts should not be issuing fatwās (Latifi, 2024). That conclusion is correct but shallow. It leaves three sharper questions unanswered. What exactly is the “authority” at stake, and is all of it the sort of thing a machine could ever acquire? Is the contrast with human reasoning really one of probability versus certainty, given that ijtihād itself classically produces probable, not certain, conclusions? And are the remedies most often proposed—especially the call to have systems simply present “all the opinions”—as innocent as they sound?

This article takes up those three questions in turn. Its central claim is that the imitation LLMs perform is genuinely deep at the level of surface propositions and genuinely empty at the level of standing: a system can approximate what a jurist would say without occupying any of the positions—epistemic, moral, communal, and covenantal—from which a jurist is entitled to say it and to be believed. To make that claim precise, the study distinguishes epistemic from deontic authority, reframes the probability question through the classical category of ẓann, and reads the isnād tradition as an early provenance protocol whose logic exposes what retrieval-based systems can and cannot fix. Methodologically, the argument proceeds by comparative and conceptual analysis: it sets the epistemic commitments of classical ijtihād against the architecture and documented behavior of current systems, drawing on recent benchmark evaluations of Islamic chatbots (Abbasi, 2025; Atif et al., 2025), the technical literature on retrieval-augmented generation (Ayala & Bechard, 2024; Song et al., 2024), and two non-sectarian ethical frameworks—the Rome Call for AI Ethics (Pontifical Academy for Life, 2020) and the G7 Hiroshima AI Process (G7, 2023)—that independently converge on the principle organizing this paper: that advanced AI should augment human judgment rather than replace it in domains where the stakes are high and the values fundamental.

The Epistemic Foundations of Ijtihād and the Digital Turn

The Classical Architecture of Ijtihād

In the classical tradition, ijtihād is not a synonym for “interpretation” in the loose modern sense. It is a demanding, credentialed effort—“the utmost exertion,” as the term’s root implies— by which a qualified scholar (mujtahid) derives a ruling (ḥukm) for a novel or contested question from the revealed sources and the disciplined machinery built to read them (Hallaq, 2009; Kamali, 2003). The competences required are specific and cumulative: command of Arabic; knowledge of the Qurʾān and its legal verses; mastery of the prophetic reports (ḥadīth) and of the sciences that grade their reliability (ʿilm al-rijāl, the appraisal of narrators); familiarity with the settled points of consensus (ijmāʿ); facility with analogical reasoning (qiyās); and an internalized sense of the higher aims of the law (maqāṣid al-sharīʿa). One does not become a mujtahid by reading widely; one is recognized as one by teachers who themselves hold the license (ijāzah) to certify, within a chain of transmission that mirrors, at the level of persons, the isnād that authenticates texts.

Three features of this architecture will matter for everything that follows, and each is routinely flattened in the AI debate. The first is that ijtihād is epistemically modest by design. Outside a small class of definitive matters (qaṭʿiyyāt)—the existence of the five daily prayers, the prohibition of intoxicants, and the like, where the evidence is decisive—the mujtahid’s conclusion is understood to be ẓann: a considered, method-governed probable opinion, not a certainty (al-

Ghazālī, 1993; Kamali, 2003). The tradition is thus not naïve about uncertainty; it is built around it. A ruling is defended not by claiming that it is certainly God’s law but by showing that it rests on sound evidence and correct method—a distinction to which this article will return, because it dissolves the tidy “certainty versus probability” contrast that so much of the literature relies on.

The second feature is that disagreement is constitutive rather than embarrassing. The coexistence of multiple recognized schools (madhāhib)—the Ḥanafī, Mālikī, Shāfiʿī, and Ḥanbalī within Sunnī Islam, the Jaʿfarī within Shīʿī Islam, and others—reflects a structural commitment to principled pluralism (ikhtilāf). This pluralism is not a menu from which anyone may order freely; it is a landscape of positions, each anchored in evidence and method, and each internally ordered by tarjīḥ, the weighing of competing views to identify the stronger within a given school. That the classical compendia record thousands of documented disagreements is not a sign of a system in disrepair but of one that took seriously both the richness of the sources and the diversity of human circumstance (Kamali, 2003).

The third feature is that the mujtahid is a bearer of obligation, not merely a producer of opinions. The scholar operates under taklīf—moral and legal accountability—understood to run all the way to the divine plane. A widely transmitted prophetic report frames the stakes vividly: the scholar who exerts genuine effort and reaches the correct conclusion earns two rewards, while the one who errs, having exerted the same sincere effort, still earns one. Whatever one makes of the theology, its practical effect is unmistakable: it attaches personal, non-transferable responsibility to the act of ruling, and it makes intellectual honesty, thoroughness, and humility not optional virtues but conditions of legitimacy. Auda (2008), restating this tradition through the language of systems theory, captures what such accountability protects—the capacity to weigh competing values, to read the priorities of a particular situation, and to exercise moral judgment under real uncertainty. That capacity, he argues, is where the maqāṣid actually live, and it is precisely the capacity that current systems are least equipped to hold.

From Print to Platform: The Mediatization of Islamic Knowledge

The scholarship on religion and media offers a sober corrective to any sense that the present moment is unprecedented in kind rather than degree. Bunt (2009) traced how fatwā websites and digital Qurʾān repositories democratized access while fragmenting the older architecture of authority; Campbell and Tsuria (2021) generalized the pattern into a theory of “digital religion,” showing that new media do not simply deliver religious content but restructure the very conditions under which religious meaning is produced, circulated, and authorized. Messick’s (1993) account of the shift from manuscript to print in Yemen, and Hirschkind’s (2006) study of the cassette sermon in Egypt, both document the same double movement: each technology expands who can reach religious knowledge and simultaneously erodes the gatekeeping of the institutions that had governed it. The LLM is continuous with this history. What is new is that earlier media externalized storage and transmission, whereas the LLM externalizes something that looks like reasoning—and it is the appearance of outsourced reasoning, rather than outsourced storage, that unsettles the classical settlement.

Islamic Chatbots and the Emergence of Algorithmic Authority

Ahmad’s (2025) survey, the most comprehensive to date, maps a fast-diversifying ecosystem and offers a taxonomy along five axes: governance (state-backed, commercial, or open- source), jurisprudential orientation (Sunnī, Shīʿī, or non-denominational), scope (Qurʾān-only, ḥadīth-focused, jurisprudential, or general), technical architecture (fine-tuned, retrieval- augmented, or prompt-engineered), and transparency (open versus proprietary). The taxonomy is useful precisely because it makes visible how much of a system’s effective doctrine is decided by unstated engineering choices—which corpus was indexed, which school’s manuals were most represented, how retrieval was tuned—rather than by any deliberative process a scholar would recognize.

It is worth stating plainly why the classical mechanisms have no structural counterpart inside these systems. Traditional authority rests on verifiable warrants: the isnād that authenticates a report, the ijāzah that certifies a scholar, the ijmāʿ that stabilizes a settled position. An LLM produces its output by estimating the most probable continuation of a text given patterns learned from a training corpus. There is no point in that process at which a narrator is graded, a license is checked, or a consensus is consulted as a consensus; there is only the statistical shadow that such things cast in the text the model happened to ingest. The result can be fluent and even, on well- trodden questions, reliably correct—yet it arrives without what one might call epistemic traceability: the ability of a user to follow the reasoning back through authenticated sources and accountable persons to the point where responsibility for it rests. The remainder of this article is, in a sense, an extended argument about what that missing traceability does and does not entail.

Empirical Evidence: How Current Systems Actually Perform

The abstract worry acquires teeth only against evidence, and the most rigorous evidence now available is discouraging in instructive ways. Atif et al. (2025), in a study presented at the AAAI/ACM Conference on AI, Ethics, and Society, built a benchmark—FiqhQA—of 960 question–answer pairs covering the four major Sunnī schools in both Arabic and English, and evaluated six models spanning closed and open weights, Arabic-focused and multilingual, large and small. Two findings bear directly on the present argument. First, performance varied sharply across languages and across schools of law: models did not treat the madhāhib even-handedly, and some were markedly better represented than others in the models’ outputs—an empirical signature of exactly the pluralism problem discussed below. Second, and more revealing, the strongest model on raw accuracy was not the strongest on abstention—the capacity to decline when it should not answer. The most accurate system was among the least willing to say “I do not know,” while other systems traded some accuracy for the discipline of silence. In a domain where a confident wrong answer about the will of God is worse than no answer at all, that trade-off is not a technicality; it is the whole game.

Abbasi’s (2025) study of Islamic inheritance law (ʿilm al-farāʾiḍ) sharpens the point from the opposite direction. Inheritance is the most computationally tractable corner of the law: shares are fixed, the arithmetic is determinate, and dedicated calculators have existed for years. If any area should be safe for a language model, it is this one. Yet Abbasi found that leading systems, while capable of impressively synthesizing and comparing Sunnī and Shīʿī rules, nonetheless generated inaccurate citations and misapplied nuanced principles, occasionally confusing the rules of different schools. The lesson generalizes uncomfortably: if models stumble where the task is essentially mathematical, their reliability in the interpretive, context-laden domains that make up most of the law—family disputes, financial ethics, bioethics—can only be lower. The empirical record, in short, does not describe a technology that is nearly ready and merely needs polish. It describes one whose failure modes are structural, and it is to that structure that the conceptual analysis now turns.

Conceptual Framework

Two Kinds of Authority: Epistemic and Deontic

The single most useful distinction for clarifying this debate is one the debate rarely draws. Authority, in matters of religious guidance, comes in two forms that ordinary language runs together. Epistemic authority is the property of being a reliable source of true propositions: an epistemic authority on inheritance is a source whose answers about shares tend to be correct. Deontic authority is different in kind. It is the standing to oblige—to issue guidance that the questioner has some reason to treat as binding on conduct, and for which the issuer answers. A muftī possesses both: reliability about the law, and the recognized standing, grounded in certification and communal acceptance, to render a fatwā that a petitioner may act upon.

Separating the two immediately reorganizes the question. Whether an LLM can attain epistemic authority is, at bottom, an empirical matter of accuracy and calibration; nothing in principle forbids a future system from becoming a highly reliable source of true legal propositions on well-structured questions, even if present systems fall short (Abbasi, 2025; Atif et al., 2025). Deontic authority is another thing entirely, because it is not a property of outputs at all. It is constituted socially and theologically—by the ijāzah that certifies, by the community that recognizes, and above all by the taklīf that renders the issuer accountable. And accountability cannot be conferred on a system that bears no obligation, feels no weight, and can incur no blame. One can meaningfully hold a muftī responsible for a ruling that wrecks a marriage or blesses an unjust transaction; there is no corresponding sense in which a model can be held responsible, because there is no bearer there to hold. This is why “algorithmic religious authority” is best understood as a real sociological phenomenon and an impossible normative one. Users do, demonstrably, treat fluent outputs as if they bound (Latifi, 2024); but nothing about the system can actually bind, because binding is an act performed by an accountable agent, and the model is not one. The danger, then, is not that machines will acquire deontic authority. It is that people will behave as though they had, transferring to an unaccountable process a deference that the tradition reserves for accountable persons.

The Two Probabilities: ẒAnn Versus Stochastic Fluency

With that distinction in hand, the field’s favorite contrast—probabilistic machine versus certain human—can be corrected. It is simply not the case that the mujtahid trades in certainties. As noted above, ijtihād by its own self-understanding yields ẓann, disciplined probable opinion, across the vast interpretive terrain that lies outside the definitive matters (al-Ghazālī, 1993; Kamali, 2003). Both the jurist and the model, that is, operate in the register of the probable. The real difference lies elsewhere, and it is twofold. First, the two probabilities range over different objects. The model estimates P(next token | preceding text): a probability over linguistic surface form, learned from how words have co- occurred in a corpus. The mujtahid’s ẓann is a probability over the ḥukm—an estimate of what God’s law requires given the weight of the evidence (adilla)—reached through a public method that can be stated, contested, and revised. These are not two grades of the same quantity; they are conjectures about entirely different things. A model that is extremely confident about the next word can be, and often is, entirely silent about the ruling, because fluency in the medium is not the same as warrant in the matter. The category error that produces the illusion of algorithmic authority is precisely the conflation of a probability-over-strings with a ẓann-over-ḥukm—mistaking the smoothness of the sentence for the soundness of the judgment it expresses. Second, the two probabilities differ in discipline and ownership. The mujtahid’s ẓann is not a private hunch; it is answerable to a method (uṣūl) that others can check, and it is owned by an agent who stands behind it under taklīf. The model’s distribution is answerable to nothing but its training data and objective, and it is owned by no one in the relevant sense. This is why the abstention finding in Atif et al. (2025) is so telling. A conscientious jurist’s awareness that a question exceeds their competence produces silence—lā adrī, “I do not know,” has always been counted among the finest words a scholar can utter. A model optimized to produce a fluent completion has no comparable internal check; its default is to continue, and continuing, in this domain, is where the harm lives. The problem is not that the model is probabilistic. It is that its probability is undisciplined by method and unowned by any agent who could be answerable for it.

Isnād as Provenance: Why Better Retrieval Solves the Wrong Problem

The most common technical response to these worries is to ground the model’s outputs in authenticated sources through retrieval-augmented generation (RAG), and the most common hope attached to that response is that grounding will, in time, restore something like traditional reliability. Reading the isnād system as an early provenance protocol shows why that hope is half- right in a way that matters. The isnād was, in effect, a distributed verification technology: every transmitted report carried a chain of named transmitters, and each link was appraised for reliability and continuity, so that the community could trace a statement back to its source and assign it a grade. It is a strikingly modern-sounding solution to a problem the AI field is now rediscovering under the banners of data lineage, provenance, and verifiable attribution—and RAG’s practice of surfacing source citations is a crude gesture in the same direction.

But here is the decisive point. The isnād authenticates transmission—that a particular person reliably reported a particular text—and it says nothing, by itself, about interpretation: what that text means for a novel case, how it interacts with competing evidence, or which of several sound readings should govern. Authentication and interpretation are different labors, performed by different competences. A perfect isnād establishes that a ḥadīth is sound; it does not tell you the ḥukm. The upshot is that even a hypothetical system with flawless “computational isnād”— one that grounded every clause in an authenticated chain—would still not be performing ijtihād. It would be an excellent transmitter, a tireless clerk of the sources, a superb muḥaddith’s assistant. It would not be a mujtahid, because the interpretive step that turns authenticated sources into a ruling is exactly the step that provenance cannot supply. This dissolves the techno-optimist’s recurring wager that enough grounding will eventually yield authority. Grounding closes the transmission gap. The interpretation gap is a different gap, and no amount of retrieval reaches across it.

Cognitive Vulnerability and the Argument From Stewardship

A further consideration turns a supposed human weakness into the ground of human authority. In the standard AI narrative, our finitude—limited memory, cognitive bias, emotional interference—is a set of deficiencies for machines to overcome. The classical tradition suggests the reverse. The mujtahid’s awareness of fallibility is not a bug to be engineered away; it is the source of the intellectual humility (tawāḍuʿ) and fear of divine accountability (taqwā) that the tradition names as prerequisites for legitimate reasoning. A theological image sharpens the point. In the Qurʾānic account of creation, the angels—perfect in obedience and, one might say, computational purity—are not granted stewardship (khilāfa); it is the flawed, mortal, learning human being who receives that trust (Qurʾān 2:30–34). Read into the domain of legal reasoning, the image suggests that the authority of the mujtahid rests not on calculative superiority, which machines will inevitably surpass, but on a capacity that is inseparable from vulnerability: the capacity for moral struggle, for growth, for repentance, and for bearing responsibility before God. A model cannot feel the weight of a fatwā on marriage, inheritance, or the permissibility of a contract; it cannot dread being wrong. The point is not sentimental. It is that some forms of knowledge are constitutively tied to the kind of being that produces them, and legal-moral judgment under accountability is one of them.

An Analogy From the Board: AlphaGo, Ten Years On

The game of Go has become an unexpectedly precise laboratory for this argument, and a decade of hindsight has made it more instructive, not less. In 2016, DeepMind’s AlphaGo defeated Lee Sedol, and its Move 37 in the second game—a play so improbable that professional commentators assumed it was a mistake—demonstrated a form of machine creativity that exceeded human expectation (Silver et al., 2016). Yet in the fourth game, Lee answered with his own Move 78, the “God’s touch,” exploiting a blind spot in the machine’s evaluation and reminding the world that human intuition retains a capacity for the unforeseen that arises from a different source than exhaustive search. The two moves, set side by side, capture the whole tension: machines can out-calculate us, and something in human understanding is not reducible to calculation.

The decade since is where the analogy earns its keep. A recent account of professional Go finds that AI has not merely raised the level of play but reshaped how players think—and not entirely for the better (MIT Technology Review, 2026). Top players now train primarily by studying and imitating engine recommendations; the strongest professionals match the machine’s preferred move markedly more often than the field average, opening sequences that once carried personal style have been standardized into near-identical lines, and players openly admit that they execute superhuman moves without grasping the principles behind them—a condition one might call epistemic limbo, in which the right move is known but the reason for it is not. The report notes both a democratization of access and a homogenization of style: more players can now reach a high level, and the styles at that level increasingly converge on the engine’s. Transposed to ijtihād, the warning is exact. A tradition whose vitality has always depended on the friction of principled disagreement, and on the slow formation that comes from wrestling with difficult texts, does not merely risk errors if it outsources reasoning to a fluent engine. It risks the homogenization of the reasoning itself—the quiet conversion of a multi-vocal, deliberative practice into the collective imitation of a single optimized voice. That is not a hypothetical extrapolation; it is what has demonstrably happened to another ancient practice under the same pressure.

Convergent Ethical Frameworks: The Rome Call and the Hiroshima Process

The concerns developed here are not parochial. They rhyme with an emerging cross- cultural consensus on AI governance. The Rome Call for AI Ethics, issued by the Pontifical Academy for Life in February 2020 and originally signed by Microsoft, IBM, the Food and Agriculture Organization, and the Italian government, sets out six principles—transparency, inclusion, responsibility, impartiality, reliability, and security and privacy—organized around the insistence that technology must serve human dignity and that AI must not be built to replace human agency (Pontifical Academy for Life, 2020). In January 2023 the Call took on a notably interreligious character when Muslim and Jewish leaders signed alongside the Vatican, making it, on its organizers’ account, the first document on which the three Abrahamic traditions had jointly agreed—a fact that lends unusual weight to its relevance for the present case. The G7 Hiroshima AI Process, launched in May 2023 and culminating in a comprehensive policy framework that December, adds an intergovernmental register, offering guiding principles and a voluntary code of conduct for organizations developing advanced systems, with emphasis on trustworthiness, transparency, and accountability (G7, 2023). Both frameworks converge, from outside the Islamic tradition, on the principle this article defends from within it: that in domains where error is costly and values are fundamental, AI should augment human capability rather than supplant human judgment.

Findings and Discussion

The Four-Dimensional Gap, Restated

Read through the distinctions above, the gap between what current systems do and what ijtihād requires resolves into four dimensions, each of which is a specific failure of the epistemic infrastructure rather than a generic shortfall in “intelligence.” The first is source authentication. Classical practice demands that every report used in reasoning be graded for the reliability of its chain and narrators. A model has no structural capacity to do this; it processes authenticated and fabricated reports alike as text, and the FiqhQA evaluation confirms that current systems frequently fail to handle reliability distinctions in the ways a competent scholar would require (Atif et al., 2025). The second is methodological transparency. A classical fatwā exposes its reasoning—naming the relevant verses and reports, specifying the school within which it operates, acknowledging alternative positions, and stating the principles that govern the derivation—so that it can be checked. A model’s output emerges from an opaque statistical process, and even when RAG surfaces citations, the selection, weighting, and synthesis of those sources remain hidden inside the architecture. Provenance of the inputs is not the same as transparency of the reasoning, and it is the latter that the tradition demands.

The third dimension is pluralistic reasoning, and it deserves the extended treatment given below, because it is where both the problem and the fashionable remedy go wrong. The fourth is deontic accountability—the dimension the epistemic/deontic distinction was designed to isolate. The mujtahid stands under taklīf, personally answerable for guidance, with the fear of consequence functioning as a discipline on honesty and care. A model occupies no such position: it cannot fear error, feel remorse for a misleading answer, or undergo the moral formation the tradition treats as inseparable from sound reasoning. This is not a philosophical flourish appended to a technical discussion; it is the practical reason the other three gaps are so hard to close. Authentication, transparency, and pluralism are, in the human case, upheld because an accountable agent is answerable for getting them right. Remove the accountable agent and you have not merely lost one feature among four; you have removed the thing that motivated the maintenance of the other three.

The Collapse of Ikhtilāf Into Monoculture

The pluralism dimension is where an apparently technical property of LLMs produces a genuinely civilizational risk. Because a language model is optimized to produce a single fluent completion, it tends to elevate one position as the answer while rendering the alternatives invisible—flattening a landscape of principled disagreement into a single output. Ahmad (2025) documents a related pathology in which systems silently blend the positions of different schools, producing a hybrid that no recognized scholar would endorse and that the user has no way to detect. The FiqhQA results give the concern empirical form: models perform unevenly across the madhāhib, systematically under-representing some (Atif et al., 2025).

The deeper danger is illuminated by work on algorithmic monoculture. Kleinberg and Raghavan (2021) prove a counterintuitive result: when many decision-makers converge on the same algorithm, the overall quality of decisions can fall even when that algorithm is more accurate than the alternatives for any individual decision-maker in isolation. Correlated reliance on a single source degrades the collective outcome, no external shock required. The implication for religious guidance is stark and specific. Even granting the optimistic premise that a chatbot reliably surfaces the single “best” opinion, mass convergence on it would still impoverish the tradition, because the epistemic health of Islamic law has never resided in everyone holding the best opinion. It has resided in the distribution of well-reasoned positions—in a community aware that sincere scholars differ, which is what keeps legal thought supple, self-correcting, and humble. A monoculture optimized for the best single answer is not a lesser version of ikhtilāf; it is the negation of the epistemic good that ikhtilāf exists to secure. The Go case is the empirical companion to the theorem: convergence on a superior engine did not just risk homogenization, it produced it (MIT Technology Review, 2026).

The Opposite Failure: The Talfīq Trap

The remedy most often proposed for monoculture is to have systems present the full range of opinions—an “algorithmic pluralism” treated in the literature as an unqualified good. It is not. Presenting every position with equal salience and no guidance on tarjīḥ solves one problem by manufacturing its opposite, and the tradition has a name for the opposite. Ikhtilāf was never a self- service buffet. A lay Muslim (ʿāmmī) is expected to follow a coherent school, or a qualified muftī’s weighing of the evidence, not to assemble a personal ruling by taking the most convenient position from each school on each question—a practice known as talfīq, “patching,” which the jurists have long viewed with suspicion precisely because it lets desire masquerade as method. A chatbot that lays out all the options and leaves the choice to the user does not restore the pluralism of the scholars; it turns every user into an unqualified self-muftī, free to shop for the answer they wanted. Modern digital fatwā-shopping is not a hypothetical; it is a documented feature of the online religious landscape (Bunt, 2009), and a system designed around naïve pluralism would industrialize it.

A defensible framework must therefore hold two commitments in tension rather than collapsing into either. Pluralism belongs at the level of the system and corpus: the full landscape of recognized schools should be represented, auditable, and never silently blended. Coherence belongs at the level of user guidance: an individual seeking to act should be anchored to a consistent school or explicitly handed off to a qualified muftī for tarjīḥ, rather than being served an undifferentiated list. Call this structured pluralism—pluralism with tarjīḥ—to distinguish it from the flat menu that current reform proposals implicitly endorse. The distinction is not pedantic. It is the difference between a system that respects the architecture of the tradition and one that dissolves it from the opposite direction, replacing a single imposed voice with a cacophony in which anything can be justified.

Fatwā Is Dialogical: The Missing Mustaftī

There is a further difficulty that operates below the level of content, at the level of what kind of act a fatwā is. A fatwā is not a context-free proposition. It is a response to a particular questioner (mustaftī) in a particular situation, and the classical tradition builds this into its method through taḥqīq al-manāṭ—the verification of how a general ruling attaches to a specific, concrete case. Jackson (1992), reading a classical manual for jurisconsults, describes what he calls the muftī’s “second education”: the practical competence to read the petitioner and the local custom (ʿurf), to notice what the question omits, and to elicit the circumstances on which the ruling actually turns. This is not ancillary to iftāʾ; on the demanding view Jackson recovers, it is half of it. The istiftāʾ—the asking—and the muftī’s duty to interrogate it are constitutive of the answer.

Language models are trained to do almost the reverse. They produce context-independent, universally phrased answers optimized to satisfy the prompt as given, and they cannot perform the crucial move of eliciting the manāṭ, because they do not know what they have not been told and are not built to ask. Recent evaluations note that model reliability degrades without careful modeling of user intent and that transparency about how an answer was produced remains largely unsolved (Atif et al., 2025). The consequence is that “chatbot fatwā” is a confusion not only at the level of epistemics but at the level of speech-act pragmatics: it takes a fundamentally dialogical, situation-bound act and simulates its output while omitting the half of the act—the reading of the questioner—that made the output a fatwā rather than a generic statement of law. Shoayb Ahmed (2024), assessing how AI bears on the muftī’s role, reaches a compatible conclusion from the institutional side: the muftī’s work is not reducible to information retrieval, and the systems now on offer approximate the retrieval while missing the discernment.

Inherited Bias and the Politics of the Corpus

The pluralism problem is compounded by biases that enter beneath the level of any Islamic fine-tuning. Abid et al. (2021) documented persistent anti-Muslim associations in a leading general-purpose model, where prompts about Muslims disproportionately produced violent completions. When such distortions live in the foundational model, domain-specific tuning and retrieval may dampen but not fully neutralize them, and they can resurface in subtler forms— framing some jurisprudential positions as more “moderate” or “mainstream” than others, and thereby importing an external normative hierarchy onto a tradition that recognizes no such ranking among its authenticated schools. This makes the composition of the training and retrieval corpus a quietly consequential act of doctrinal politics. Whichever school’s manuals dominate the index will dominate the outputs; whichever traditions exist mainly as unstructured scans rather than clean, machine-readable text will be systematically underserved. Ahmad (2025) accordingly argues that developers must audit systems for inherited stereotype, securitized framing, and bias leakage, and must invest in curated corpora, targeted evaluation, and ongoing monitoring. The broader point is that there is no neutral corpus. Every choice about what to include, in what proportion, and in what condition is a choice about whose voice the system will amplify—and in a field defined by pluralism, such choices cannot be left implicit in an engineering pipeline.

The Limits of the Technical Safeguards

Retrieval-augmented generation is the leading safeguard, and it is worth crediting what it achieves before marking what it cannot. Kadous (2024), reflecting candidly on building a practical

Islamic assistant, reports that grounding answers in authenticated texts made them substantially more accurate than ungrounded generation, that validating the system took at least as much effort as building it, and that even after seminary-style testing the validation still felt insufficient. That last admission from a serious practitioner is more probative than any external critique. The technical literature explains why the residual worry is well founded: Ayala and Bechard (2024) show that RAG reduces but does not eliminate hallucination, with models liable to be distracted by irrelevant retrieved passages or to fall back on parametric memory for reasons that remain poorly understood, and Song et al. (2024) demonstrate that even sophisticated hallucination-aware tuning achieves reduction rather than elimination—evidence that fabrication is a structural property of current transformer architectures, not a bug awaiting a patch. In most domains a residual hallucination rate is a manageable cost. In a domain where a fabricated attribution to the Prophet or a misstatement of a school’s position carries spiritual and social consequences, the acceptable threshold is far lower, and “reduced but not eliminated” is not the same as “safe.” The appropriate institutional responses, accordingly, have been structural rather than merely algorithmic: fatwā services that answer only from pre-verified rulings and escalate anything beyond their scope to human scholars, and multilingual guidance systems, such as the assistant robot deployed at the Grand Mosque, that are scoped to information rather than adjudication (Ahmad, 2025). These designs concede, in their architecture, the argument of this paper: that the safe use of the technology depends on keeping the binding act in human hands.

Toward a Framework: Structured Pluralism Under Human Authority

The analysis converges on a constructive proposal. LLMs should enter the Islamic legal ecosystem as research infrastructure that augments scholarly labor, never as autonomous agents that replace it—a conclusion that aligns the classical tradition with the Rome Call and the

Hiroshima Process alike (G7, 2023; Pontifical Academy for Life, 2020). The framework below specifies five commitments; each is a sharpened, and in places corrected, version of proposals already circulating in the literature, reformulated in light of the distinctions this article has drawn.

First, deontic reservation. The binding act of iftāʾ must remain with accountable human scholars. This is not a limitation to be relaxed as models improve, because the thing being reserved—deontic authority—is not a function of model accuracy but of accountability, certification, and communal recognition, none of which a system can hold however reliable it becomes. Systems should be positioned, in interface and in framing, as assistants that prepare the ground for a scholar’s judgment, and users should be able to tell at a glance whether they are receiving educational information or a ruling. This reframes the familiar call for “human oversight” by explaining why the human is indispensable: not as a fallible checker of a superior machine, but as the only entity in the loop capable of bearing the obligation that a ruling incurs.

Second, structured pluralism with tarjīḥ. Systems must represent the full landscape of recognized schools at the level of corpus and retrieval, label every position by its jurisprudential provenance, and never silently blend traditions—while, at the level of user guidance, anchoring an individual to a coherent school or directing them to a qualified muftī rather than serving an undifferentiated menu. This holds the two failure modes apart at once: it forecloses the monoculture that a single optimized answer produces and the talfīq that a flat list of options invites. It is the article’s central refinement of the existing “algorithmic pluralism” proposal, which, taken literally, would cure one disease by inducing another.

Third, provenance to the standard of isnād, and transparency beyond it. Following the logic of the isnād tradition, every substantive claim should be traceable to an authenticated source, and systems should disclose their architectures, corpora, and moderation protocols (Ahmad, 2025).

But provenance of inputs must not be mistaken for transparency of reasoning. Systems should therefore also expose, as far as the technology allows, how sources were selected and weighed, and should communicate their known limitations—including documented hallucination behavior and the boundaries of their training data—so that a user, and a supervising scholar, can see not only what was cited but how the answer was assembled. This is where the tradition asks more of these systems than provenance alone can give.

Fourth, context elicitation and escalation. Because a fatwā turns on taḥqīq al-manāṭ, systems must be built to notice what they have not been told. Rather than answering context- sensitive questions as though they were context-free, a well-designed system should surface the circumstances on which a ruling depends, flag when a question exceeds what it can responsibly handle, and route matters of serious personal consequence—marriage, divorce, inheritance, contested transactions—to a qualified human. Modeled on services that already escalate beyond their scope (Ahmad, 2025), this commitment operationalizes the abstention discipline that current systems most conspicuously lack (Atif et al., 2025), and it treats “I cannot answer this; consult a scholar” as a feature rather than a failure.

Fifth, governance and collective ijtihād. No system should operate without the standing supervision of a recognized board of scholars representing diverse schools, charged with reviewing outputs, catching systematic error, and recalibrating over time. This institutionalizes, in a new medium, the tradition’s own preference for collective over solitary reasoning (ijtihād jamāʿī), and it connects the technical enterprise to the bodies—national fatwā councils, established academies—whose authority the systems must serve rather than usurp. The scholarly community is already moving in this direction: studies of AI in Islamic finance find scholars simultaneously alert to AI’s analytical value and to the risk of automating sacred decision-making (Priantina et al., 2025), and this dual awareness, not uncritical adoption or reflexive rejection, is the disposition the framework is meant to support.

Implementation, Language, and the Unfinished Questions

Two practical fault lines will shape whether such a framework can be realized. The first is linguistic. Islam is lived in Arabic, Urdu, Bahasa Indonesia, Turkish, Persian, Hausa, and dozens of other languages, yet the chatbot landscape remains heavily English-centric (Ahmad, 2025), and the FiqhQA results show meaningful performance gaps between languages even within a single system (Atif et al., 2025). A tradition served mainly in a language most of its adherents do not use for religious life is not being democratized so much as narrowed, and robust natural-language capabilities for classical Arabic and the major vernaculars are a precondition, not an enhancement. The second fault line is the corpus itself. As the empirical studies repeatedly show, models are only as balanced as the material they draw on, and much of the Islamic textual heritage still exists as unstructured scans rather than clean, annotated, machine-readable text (Abbasi, 2025; Ahmad, 2025). Building and curating that infrastructure—structured fatwā databases keyed to school, annotated primary texts, digitized manuscripts—is the unglamorous work on which everything else depends, and its uneven state is currently a silent source of bias that boosts dominant positions and marginalizes minority ones. Finally, regulation remains largely unwritten: the Rome Call and the Hiroshima Process offer broad principles but no domain-specific guidance for religious AI (G7, 2023; Pontifical Academy for Life, 2020), and developing such guidance will require sustained collaboration among fatwā institutions, technologists, and ethicists rather than the importation of a generic AI policy onto a domain whose values it was never designed to protect.

Conclusion

The transition from human labor to machine logic in the production of religious guidance is real, and it is neither the salvation its boosters imply nor the catastrophe its critics fear. This article has argued that seeing it clearly requires better tools than the debate has so far used. Once authority is split into its epistemic and deontic components, the core claim states itself: an LLM may in principle become a reliable source of true legal propositions, but it can never acquire the standing to oblige, because that standing is constituted by an accountability no system can bear. Once the probability question is reframed through ẓann, the contrast with human reasoning stops being the crude “certain human, uncertain machine” and becomes the precise one that matters— two conjectures over different objects, one disciplined by method and owned by an accountable agent, the other neither. And once the isnād is read as a provenance protocol, the limits of the leading technical fix come into focus: retrieval can close the transmission gap and leave the interpretation gap exactly where it was.

These distinctions also expose a danger the reform literature has missed. The intuitive cure for algorithmic monoculture—present every opinion—does not restore the pluralism of the scholars; it manufactures the anarchic talfīq the tradition took pains to prevent. A defensible path therefore runs between two flattenings: the single imposed voice of the monoculture and the undifferentiated menu of naïve pluralism. Structured pluralism under human authority—full representation at the level of corpus, coherent guidance at the level of the individual, provenance to the standard of isnād, elicitation of the manāṭ, and the binding act reserved for accountable scholars—is offered here not as a finished blueprint but as a living principle, to be refined through exactly the kind of sustained consultation (shūrā) the tradition prescribes for communal affairs (Qurʾān 42:38).

The wider significance reaches past the Islamic case. As AI moves into every domain that involves normative judgment and moral responsibility—other religious traditions, professional ethics, the law itself—the questions raised here acquire general force. The Go players who now execute superhuman moves they cannot explain (MIT Technology Review, 2026) are a parable for any practice tempted to outsource its reasoning to a fluent engine: the moves may improve while the understanding, and the community of understanding, quietly hollow out. The deepest lesson of this study is that some forms of human knowledge are not separable from the human condition that produces them—from our finitude, our answerability to one another, and, in the tradition’s own terms, our accountability before the Divine. The question was never whether a machine can perform ijtihād. It is whether the output of a machine, however fluent and however accurate in its surface details, can carry the moral weight that the tradition demands of anyone who presumes to speak about the will of God. It cannot, and the task ahead is to build systems that know it.

Statements

Conflicts of Interest

Author Note The author reports no conflicts of interest. All chatbot systems and benchmark results discussed reflect the state of a fast-moving field as of early 2026 and should be read as a snapshot rather than a settled record.

References

Abbasi, M. Z. (2025). AI as the new jurist: Can generative artificial intelligence master Islamic inheritance law? SSRN. https://doi.org/10.2139/ssrn.5658330

Abid, A., Farooqi, M., & Zou, J. (2021). Persistent anti-Muslim bias in large language models. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, 298–306. https://doi.org/10.1145/3461702.3462624

Ahmad, M. A. (2025). Islamic chatbots in the age of large language models [Paper presentation]. 5th Muslims in Machine Learning Workshop, NeurIPS 2025. arXiv. https://arxiv.org/abs/2601.06092

Ahmed, S. (2024). Fatwā and AI: A literature-based assessment on how AI impacts the role of the muftī. Journal for Islamic Studies, 42(1–2). https://doi.org/10.25159/2957-9163/14400

al-Ghazālī, A. Ḥ. (1993). Al-Mustaṣfā min ʿilm al-uṣūl (Vol. 2). Dār al-Kutub al-ʿIlmiyya.

al-Zuḥaylī, W. (2009). Uṣūl al-fiqh al-Islāmī (Vol. 2). Dār al-Fikr.

Atif, F., Askarbekuly, N., Darwish, K., & Choudhury, M. (2025). Sacred or synthetic? Evaluating LLM reliability and abstention for religious questions. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(1), 217–226. https://doi.org/10.1609/aies.v8i1.36543

Auda, J. (2008). Maqāṣid al-sharīʿah as philosophy of Islamic law: A systems approach. International Institute of Islamic Thought.

Ayala, O., & Bechard, P. (2024). Reducing hallucination in structured outputs via retrieval- augmented generation. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Vol. 6: Industry Track), 228–238. https://doi.org/10.18653/v1/2024.naacl-industry.19

Bunt, G. R. (2009). iMuslims: Rewiring the house of Islam. University of North Carolina Press.

Campbell, H. A., & Tsuria, R. (Eds.). (2021). Digital religion: Understanding religious practice in digital media (2nd ed.). Routledge.

G7. (2023). Hiroshima process international guiding principles for organizations developing advanced AI systems and comprehensive policy framework. https://www.soumu.go.jp/hiroshimaaiprocess/en/index.html

Hallaq, W. B. (2009). Sharīʿa: Theory, practice, transformations. Cambridge University Press.

Hirschkind, C. (2006). The ethical soundscape: Cassette sermons and Islamic counterpublics. Columbia University Press.

Jackson, S. A. (1992). The second education of the muftī: Notes on Shihāb al-Dīn al-Qarāfī’s tips to the jurisconsult. The Muslim World, 82(3–4), 201–217. https://doi.org/10.1111/j.1478- 1913.1992.tb03553.x

Kadous, W. (2024, February 29). Lessons learned building a practical AI assistant [Blog post]. Medium. https://waleedk.medium.com/what-ive-learned-building-a-real-ai-assistant- 5d520b0e6a46

Kamali, M. H. (2003). Principles of Islamic jurisprudence (3rd rev. ed.). Islamic Texts Society.

Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), Article e2018340118. https://doi.org/10.1073/pnas.2018340118

Latifi, H. (2024). Challenges of using artificial intelligence in the process of Shiʿi ijtihād. Religions, 15(5), Article 541. https://doi.org/10.3390/rel15050541

Messick, B. (1993). The calligraphic state: Textual domination and history in a Muslim society. University of California Press.

MIT Technology Review. (2026, February 27). AI is rewiring how the world’s best Go players think. https://www.technologyreview.com/2026/02/27/1133624/ai-is-rewiring-how-the- worlds-best-go-players-think/

Pontifical Academy for Life. (2020, February 28). Rome call for AI ethics. https://www.romecall.org/the-call/

Priantina, A., Uula, M. M., Aufa, & Herindar, E. (2025). AI in fatwa formulation: Transforming Sharia-compliant finance. Journal of Central Banking Law and Institutions, 4(3), 595–634. https://doi.org/10.21098/jcli.v4i3.446

Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. https://doi.org/10.1038/nature16961

Song, J., Wang, X., Zhu, J., Wu, Y., Cheng, X., Zhong, R., & Niu, C. (2024). RAG-HAT: A hallucination-aware tuning pipeline for LLM in retrieval-augmented generation. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, 1548–1558. https://doi.org/10.18653/v1/2024.emnlp- industry.113