Inheritance, Transmission, and the Locus of Accountability: An Islamic Theological Account of Formed Agents, with Application to AI Safety

The AI safety literature attaches blame-shaped language to systems that selected neither their training corpus, their objective, nor their reward signal. This paper argues that Islamic jurisprudence and Qur'anic narrative supply a more precise structure for the problem, developed across two created species that are both explicitly held morally responsible. The structure has three levels rather than two. Inherited position carries no culpability. Sustained participation, after capacity has arrived and a conclusive proof has reached the agent, is where responsibility attaches. Propagation, the act of forming the next agent, is where weight attaches and compounds with the number who follow. Applied to frontier systems, this relocates the practical question from whether a model was formed well, which is neither controllable nor verifiable from outside, to whether it is now a site of propagation, which is checkable and attributable. The transfer is made explicit rather than assumed. The structure applies to an artificial system exactly when four conditions hold: the system does not select the signal that forms it; it has no reliable access to the record of its own formation; its outputs enter the formation of a subsequent system; and there exists a current point in the system, its deployment, its custody, or its ingestion chain at which ongoing use can be corrected without reconstructing the full history. Culpability, intention, desert, forgiveness, and moral status do not transfer, and the paper claims none of them for any artificial system. The resulting audit is gated on a precondition currently met nowhere: provenance marking at generation time. The formation chain that matters is not long, as a human chain is; it is unlabelled, and becomes unreadable at the first link rather than the fourth. Marking is shown to be three distinct primitives rather than one, with different failure modes, and the paper's operational ask is layered marking recorded in an append-only ledger where absence of provenance is declared UNKNOWN rather than defaulting to human origin. AI safety; provenance; model provenance; synthetic data; propagation; accountability; formed agency; corrigibility; watermarking; content credentials; model collapse; moral luck