Dr Anye Nyamnjoh
Senior Research Officer, EthicsLab
22 Apr 2026

Original site of this blog post: An Output Is Not a Decision: Ifá, AI, and the Ethical Labour of Interpretation | EthicsLab

In our recent EthicsLab webinar, Uyiosa Omoregie read the West African Ifá system as a form of structured computation. The most productive thing his provocation offers AI ethics, I want to suggest, is not an origin story but a question: when does an output become a decision, and who is authorised to interpret it?

Omoregie’s work is easy to misread in two opposite but equally tempting ways: as a claim of equivalence, that Ifá is AI, or as a claim to historiographical prestige, that Africans had binary logic first. Either reading turns the discussion into an argument about precedence, which is precisely not the point. He neither claims that a Yoruba divination system is a large language model, nor that the history of computing has a single African headwater. What he asks is narrower and more provocative: if we widen our definition of computation to include a rule-governed symbolic practice that produces structured outcomes and depends on trained interpretation, what might we notice about our own machines that we had stopped noticing?

This is the thread worth pulling. Systems become authoritative through social arrangements, not only through technical procedures. An output does not automatically carry the status of a decision. Its authority is made, granted, withheld, contested, by the people and institutions around it. Read this way, Ifá is less a precursor to our machines than a way of seeing them more clearly.

Interpretation as part of the system, not a step after it
The structural feature that makes Ifá a generative reference point is that the outcome is never allowed to stand alone. When a babaláwo casts the opele, the resulting odù, one of 256 binary configurations, each mapped to a vast oral corpus of verses, opens a reference point. Rather than pronouncing a verdict, meaning emerges through dialogue: between the practitioner, the client, and the community whose recognised members are authorised to interpret. The same odù that speaks to a farmer worried about a harvest speaks differently to a parent worried about a child. Interpretation is constitutive of computation and not just a downstream courtesy added to it.

This is a narrow claim, and worth keeping narrow. It is not that Ifá is “more ethical” in some sweeping sense. Instead, Ifá foregrounds interpretation and responsibility as the conditions under which an output can responsibly guide action. One participant in the discussion pushed this further, noting that such systems are not only predictive but worldmaking: they do not merely forecast a future, they help shape who a person should become. That observation sharpens the stakes considerably. If a system’s outputs participate in making worlds, then the labour of interpreting those outputs is exactly the place where that worldmaking should happen in the open: carried by people who can be named and held to account, rather than settled silently by the bare fact of a result.

Digital divination?
Turn that lens on contemporary AI and the worry is hardly mysticism. Calling a model a “digital diviner” can sound sensational, but, used carefully, the phrase names a precise pattern. In some settings an output starts to be treated less like one input among many and more like an authorising signal, something that settles questions, justifies action, and lowers the need for explanation. The live question, the one I think Omoregie’s provocation presses hardest, is therefore one of design and institutional arrangement: to what extent are our systems built, deployed, embedded, and trusted in ways that minimise interpretation, so that the output arrives already wearing the costume of an answer?

This neither describes all AI systems, nor is it inevitable. But it is observable in certain applications where three conditions converge: high stakes under time pressure, where people want a decision quickly; technical complexity, which makes outputs feel difficult to challenge; and institutional exposure, where a contestable decision is easier to defend if it can be pinned on the system. Where they hold, an output begins to function like an oracle, because it has been socially positioned as an authority, as opposed to some inherent mysticism. Rather than distrusting models, the corrective is to refuse a premise that has stealthily become a default: that an output is self-justifying. It is not! An output is not yet a decision.

Consider clinical decision support, like an early-warning score that flags a patient as deteriorating, or a model estimating risk. The ethical question is not only whether the model is accurate on average. It is how its output becomes actionable on the ward. A well-documented hazard here is automation bias: in a systematic review of clinical decision support, Goddard and colleagues found that clinicians can over-rely on automated advice and miss errors the system itself introduces, with the effect strongest under precisely the conditions named above (time pressure and task complexity).

What does the damage is rarely the score itself, but the asymmetry that flourishes around it. Following the score requires no explanation, while deviating from it must be justified. That asymmetry relocates authority onto the output. And when something goes wrong, it relocates responsibility too: the clinician who was nominally “in the loop” becomes what Madeleine Elish calls a moral crumple zone, the nearest human, absorbing blame for a system over which they had limited real control, while the integrity of the technology is preserved. The output had become the decision and the human was left holding the consequence.

That is the precise inversion of the Ifá arrangement, in which the interpreter is authorised and resourced to interpret and the community shares in the judgement. The takeaway is not to import Ifá into the hospital, but to attend to the conditions Ifá makes visible, keeping interpretation central, shared, explicit, and contestable. Concretely, in the clinic, that can mean requiring documented clinical reasoning for decisions that incorporate but do not defer to the score; making justification symmetrical, so that following the model and departing from it both call for stated reasons; establishing review for contested outcomes; and ensuring patients can ask how a decision about them was reached.

Anticipating the obvious objections

Two challenges are worth meeting directly:
Is “a human in the loop” not already the answer? Not on its own. Ben Green’s analysis of human-oversight requirements shows that mandating a person to review or sign off frequently fails to deliver accountability, and can manufacture the very rubber stamp it was meant to prevent. The point here is qualitative rather than procedural: interpretation has to be authorised, resourced, and genuinely able to overturn the output, and not just a signature collected to launder it.

Does mandated interpretation not just add friction, or invite rubber-stamping? It can, which is exactly why the three adjectives carry weight. Interpretation that is merely required becomes ceremonial. Interpretation that is shared (carried by recognised, competent people), explicit (surfacing reasons rather than deferring to a number), and contestable (open to challenge by those it affects) is friction of the productive kind, the kind that makes reasons visible and decisions answerable.

A cautious takeaway
AI ethics often asks about the model: is it accurate, is it fair, is it transparent? Omoregie’s provocation, taken at its strongest, asks instead about the social conditions under which an output becomes authoritative. If the image of AI as a “digital diviner” is worth keeping, it is worth keeping only as a discipline, pointing not to mystery, but to the work of keeping outputs interpretable, contestable, and accountable, rather than letting them become self-justifying pronouncements that settle questions simply by being produced. Uyiosa Omoregie offers a compelling reminder that what modern computation might learn from Ifá is a critical posture toward its own authority.

Watch the Webinar here: