The Shepherd and the Algorithm

Something goes wrong. An algorithm denies a loan. A facial recognition system flags the wrong person. A diagnostic tool recommends a treatment that makes the patient worse.

The person harmed looks for someone to hold accountable. What they find is a chain of humans, each gesturing past themselves. The engineer says: I built the model, but I didn’t deploy it in this context. The company says: we provided the tool, but the operator configured it. The operator says: I followed the vendor’s guidelines. The end user says: I just pressed a button.

Each defense is individually plausible. Collectively, they produce a world in which harm was done and no one is responsible.

This is not a failure of individual character. It is a structural feature of how these systems work. The opacity of the machine — the fact that its decision-making process is often unexplainable even to its creators — generates genuine ambiguity in the causal chain. And ambiguity in causal chains, in any legal framework built around proximate cause, functions as a defense. Not because anyone designed it that way. Because the architecture of the technology naturally diffuses accountability across enough actors that no single one can be pinned.

Yesterday in Nigeria, an AI-generated image of a state governor bowing before the Sultan of Sokoto spread across WhatsApp and Facebook, weaponizing the country’s Muslim-Christian fault line in a region already shattered by unresolved killings. Who generated the image? Who trained the model that made it possible? Who shared it? An investigation is underway. The accountability is already dissolving.

The secular regulatory response has been to construct new liability frameworks after the fact — risk classifications, product liability proposals, criminal investigations into AI companies. These are serious efforts. They are also, structurally, attempts to trace a causal chain backward through layers of technical and organizational complexity that were not designed to be traced.

The Islamic legal tradition begins from the opposite end. It does not start with who caused the harm. It starts with who accepted the trust.

In Surat al-Ahzab, Allah describes a moment before human history:

إنا عرضنا الأمانة على السماوات والأرض والجبال فأبين أن يحملنها وأشفقن منها وحملها الإنسان إنه كان ظلوما جهولا

“Indeed, We offered the Trust to the heavens and the earth and the mountains, and they declined to bear it and feared it; but man undertook to bear it. Indeed, he was unjust and ignorant.” (33:72)

The amanah is a weight the heavens themselves refused. The human being who accepted it is described as zaluman jahula — prone to injustice and ignorance. Not as condemnation, but as description. The trust was not given to a being qualified to bear it perfectly. It was accepted by one who would struggle, fail, and still be held accountable for the acceptance.

When a person builds or deploys a system that makes decisions affecting other people — their creditworthiness, their liberty, their medical treatment — they have accepted a trust. The system itself has no share in this. It has no moral capacity (taklif), no rational agency, no genuine choice. The amanah cannot be transferred to it any more than it could be transferred to the mountains.

The Prophet, peace be upon him, gave this principle its operational structure in the hadith narrated by Ibn Umar:

كلكم راع وكلكم مسئول عن رعيته والأمير راع والرجل راع على أهل بيته والمرأة راعية على بيت زوجها وولده فكلكم راع وكلكم مسئول عن رعيته

“Each of you is a shepherd, and each of you is responsible for his flock. The leader is a shepherd. The man is a shepherd over his household. The woman is a shepherdess over her husband’s house and children. Each of you is a shepherd, and each of you is responsible for his flock.” (Bukhari and Muslim)

Two features of this model speak directly to the accountability void that AI creates.

First, responsibility is layered and non-transferable. The ruler’s obligation to his subjects does not dissolve when he delegates to a governor. The governor does not escape by pointing to the ruler’s instructions. Each layer holds its own stewardship. Applied to the AI deployment chain: the executive who approved the product, the engineer who built the model, the product manager who selected the use case — each is a shepherd over their scope. The “algorithm decided” defense is not available. You are the shepherd. The algorithm is an instrument of your shepherding.

Second, the hadith does not require the shepherd to know every member of the flock personally. The ruler is responsible for subjects he will never meet. The obligation is not comprehensive knowledge of every outcome — it is active stewardship within your scope. One common defense of AI deployers is: the system makes millions of decisions; no human can review them all. The shepherd model does not ask you to review them all. It asks whether you accepted responsibility for the domain and exercised care within it.

In Hanbali fiqh, this principle finds legal specificity. The Zad al-Mustaqni states that an agent — a wakil — is a trustee who bears no liability for what is damaged in his possession without negligence:

والوكيل أمين لا يضمن ما تلف بيده بلا تفريط

But the protection holds only in the absence of negligence — tafrit. The moment negligence enters, the liability attaches. An AI system cannot be negligent. It has no will, no carelessness, no capacity for tafrit. It is an instrument. The negligence question, therefore, always refers back to the human principal: did they deploy with due diligence? Did they understand the system’s limitations? Did they monitor its consequences?

The deepest challenge AI poses for accountability is the “I didn’t know” defense. I didn’t know the model was biased. I didn’t know the training data contained personal information. I didn’t know the output would be used this way.

The Quran addresses this:

ولا تقف ما ليس لك به علم إن السمع والبصر والفؤاد كل أولئك كان عنه مسئولا

“And do not pursue that of which you have no knowledge. Indeed, the hearing, the sight, and the heart — about all those will be questioned.” (17:36)

Ignorance is not neutral here. It is itself something you will be questioned about. The verse does not say: if you do not know, you are excused. It says: the faculties you failed to employ — your hearing, your sight, your understanding — will testify about the failure.

In Bulugh al-Maram, Ibn Hajar records a hadith that applies this to professional negligence:

من تطبب ولم يكن بالطب معروفا فأصاب نفسا فما دونها فهو ضامن

“Whoever practices medicine without being known for medicine, and causes harm to a person or less, is liable.” (Narrated through Amr ibn Shu’ayb; reported by Abu Dawud, al-Nasa’i, al-Daraqutni; authenticated by al-Hakim)

The Arabic is damin — the person bears the guarantee. If you operate in a domain where you lack qualification, and harm results, the liability is yours. A hospital that deploys a diagnostic AI without the clinical expertise to evaluate its outputs is practicing medicine through an instrument it does not understand. A company that deploys a hiring algorithm without comprehending its discriminatory patterns is making consequential human decisions without qualification. The tool is their instrument. The guarantee is theirs.

One further principle completes the framework. The Prophet, peace be upon him, said:

الخراج بالضمان

“Profit is with liability.” (Narrated by Aisha; reported by al-Tirmidhi, Ibn Khuzayma, Ibn Hibban, and al-Hakim, who authenticated it)

The rule is structural: whoever profits from a transaction bears its risk. You cannot extract the revenue an AI system generates while disclaiming responsibility for its harms. The benefit and the liability are indivisible. A company valued on the strength of its AI products cannot simultaneously argue that it bears no responsibility for what those products do. Al-khiraj bil-daman — the profit and the guarantee travel together.

This framework does not resolve every specific case. It does not calculate the precise apportionment of liability between the programmer, the deployer, and the operator in a particular AI failure. That is the work of applied legal reasoning — ijtihad — which demands domain expertise the classical scholars did not have and contemporary scholars have not yet fully developed.

But it offers what the current discourse lacks: a structure in which the opacity of a system does not reduce responsibility but increases it. The harder a system is to understand, the heavier the obligation to investigate before deploying it — because ignorance is a liability, not an exemption. The more profitable a system, the more inescapable the guarantee — because benefit and risk cannot be severed. The longer the chain of delegation, the more each link must answer for its own stewardship — because the shepherd’s obligation does not dissolve in the size of the flock.

The question “who is responsible when the machine decides?” has a clear answer in this tradition. The machine does not decide. It processes. It executes. It optimizes. It does none of this with moral awareness, choice, or capacity to bear consequence. The decision — and therefore the responsibility — belongs entirely to the humans who built it, deployed it, and profited from it.

The mountains saw the weight of the trust and refused it. Man picked it up. It is his.