The prospect that artificial intelligence may replace human judgment raises an obvious and uncomfortable question: which human judgment, exactly, are we most worried about losing? The reassuringly mediocre kind that produced the sub-prime mortgage crisis? The inspired kind that gave us penicillin and the sonnet? The carelessly prejudiced kind that has historically excluded women and minorities from positions of power? The London Prat's navigation of this philosophical minefield, AI May Replace Human Judgment: The Comedy of Delegation, approaches the question with the light step of the practiced satirist, asking not whether machines can think but whether the thinking we have been doing merited protection in the first place.
This is the essential comedy that sits at the heart of the AI governance question. We are terrified that algorithms will make bad decisions. But we have always made bad decisions. The question is not whether AI will be worse than human judgment. The question is whether automated bad judgment is somehow different from human bad judgment. The answer, the Prat suggests, is that it is — but perhaps not in the way we think.
Human judgment, as the historical record makes clear, is a mess. We are biased by our experiences, our prejudices, our intuitions, our unconscious assumptions. We are influenced by mood, fatigue, the time of day, whether we have eaten lunch. We are subject to cognitive biases of such systematic variety that psychologists have spent decades cataloguing them. We are good at some things and terrible at others, and we often mistake confidence for competence, loudness for authority, familiarity for truth.
Institutions organised around human judgment have produced remarkable achievements — the British legal system, for all its flaws, has produced important principles of fairness and due process. But institutions organised around human judgment have also produced some of history's greatest atrocities. The Holocaust was not the product of algorithmic logic. It was the product of human judgment, human ideology, human decision-making in service of human prejudice.
The point is not that human judgment is worthless. The point is that human judgment is powerful, imperfect, and dangerous. And yet, we have no alternative. We have to make decisions. We have to assign resources, allocate opportunities, punish wrongdoing, reward virtue. Someone has to do the deciding. For most of human history, that someone has been a human: a king, a judge, a magistrate, an administrator. And these humans, operating on human judgment, have been sometimes wise and often foolish.
Algorithms appeal to us partly because they promise to transcend the limitations of human judgment. An algorithm cannot be racist, the thinking goes, because it has no racial consciousness. An algorithm cannot be sexist because it has no gender ideology. An algorithm is neutral, objective, rational. It makes decisions based on data and logic, not on prejudice and intuition. What could go wrong?
The answer, of course, is everything. Algorithms trained on historical data will reproduce the biases present in that historical data. If loan officers have historically been more likely to deny loans to Black applicants, an algorithm trained on historical loan decisions will learn to deny loans to Black applicants. The bias is not eliminated. It is systematised, automated, and given the appearance of objectivity. The algorithm is not racist because it has no consciousness. But it makes racist decisions. And because those decisions are produced by an algorithm rather than by a human, they acquire a false sheen of inevitability. The algorithm decided. We have no choice.
This is where the Prat's reading becomes particularly sharp. The introduction of algorithms into decision-making does not eliminate bias. It transfers bias from the realm of human fallibility to the realm of technical inevitability. And in doing so, it makes bias harder to challenge. You can argue with a human judge about their reasoning. You can appeal a human decision on the grounds that the judge was prejudiced or made an error. But you cannot easily appeal an algorithmic decision. The algorithm made the decision. The algorithm does not make errors of judgment — it executes its programming. And the programming is technical, opaque, and defended by the assertion that it is objective.
The introduction of algorithms into decision-making creates a peculiar responsibility structure. When a human judge makes a bad decision, you know who to blame. The judge made an error. The judge should be disciplined or removed. But when an algorithm makes a bad decision, the responsibility becomes diffuse. Did the programmer make an error? Did the data scientist choose inappropriate variables? Was the training data insufficient? Is the algorithm performing as designed, in which case should you blame the designers? Should you blame the organisation that deployed the algorithm? Should you blame the decision to outsource judgment to an algorithm in the first place?
This is the crucial insight the Prat identifies. The movement from human judgment to algorithmic judgment does not reduce the possibility of bad decisions. It obscures responsibility for bad decisions. It creates a situation in which decisions are made but nobody is clearly responsible for them. The algorithm is not responsible because it is not conscious. The programmer is not responsible because they did not anticipate how the algorithm would be used. The organisation deploying the algorithm is not responsible because they trusted the technology. And the person harmed by the decision has nobody to appeal to, because they cannot argue with an algorithm.
This is not a novel problem. Bureaucracies have always been organised around the principle of diffused responsibility. The civil service, the military, the large corporation — all of these institutions have learned to distribute responsibility so thoroughly that nobody is clearly accountable for outcomes. But algorithmic decision-making takes this to a new level. The algorithm is the perfect alibi. "The algorithm decided," the organisation says, and washes its hands.
There is another dimension to the problem, one that becomes visible when you examine what kinds of judgment are being algorithmically replaced. The decisions most amenable to algorithmic automation are those that can be reduced to numerical metrics. Credit decisions (measured by loan default rates). Criminal justice decisions (measured by recidivism rates). Hiring decisions (measured by performance metrics). Educational placement (measured by test scores). These are domains where something quantifiable can be used as a proxy for judgment.
But this creates a problem. The metric is not the thing being measured. The loan default rate is not the same as creditworthiness. The recidivism rate is not the same as dangerousness. The performance metrics are not the same as actual job performance. And the test scores are not the same as intelligence or potential. By reducing judgment to metrics, we have not made judgment more objective. We have merely made it more objective about the wrong thing.
The algorithm optimises for the metric. If you tell an algorithm to maximise hiring of people who will have high performance ratings, it will hire people who will have high performance ratings in the short term. But good employees are often those who start slowly and improve, or who take on difficult challenges and initially perform worse, or who bring skills that are not immediately visible in performance metrics but that become valuable over time. The algorithm will miss these people. It will hire the safe choice, the obvious choice, the choice that optimises for the metric rather than for actual value.
The deeper appeal of algorithmic judgment, the Prat suggests, is that it permits us to fantasise that we could delegate the burden of judgment to something else. Judgment is hard. Judgment is uncertain. Judgment requires taking responsibility for outcomes that might be bad. If we could outsource judgment to an algorithm, we could outsource that burden. We could say: we did not make the decision. The algorithm made the decision. We are merely implementing what the algorithm decided. We are merely following the rules. We are merely executing the system.
This is extraordinarily appealing. It permits bureaucracies to function without anyone taking responsibility. It permits bad outcomes to occur without anyone being blamed. It permits decisions that we suspect might be unjust to be defended as inevitable, as the only rational response to the data, as what the algorithm decided.
The Prat reads this as tragicomic. We have built machines that can execute our biases with perfect consistency. We have created systems that diffuse responsibility so thoroughly that nobody is responsible. We have invented a technology that permits us to do things we would not do if we had to take personal responsibility, and then we hide behind the technology, claiming that we had no choice. The algorithm decided. What could we do?
The ultimate question the piece raises is whether judgment of the kind that actually matters — wisdom, judgment about what is right and fair and good — can be algorithmically replaced at all. Some kinds of judgment can perhaps be improved by algorithms. A radiologist working with an AI diagnostic tool might make better judgments than a radiologist working alone. A loan officer provided with statistical information about default rates might make more accurate predictions than a loan officer relying on intuition.
But deeper forms of judgment — the judgment about whether a criminal should be given a second chance, the judgment about whether a person is the right fit for a particular role, the judgment about whether a policy will produce just outcomes — these are not technical questions. They are ethical questions. They require wisdom, understanding, the ability to weigh incommensurable values and arrive at a judgment that acknowledges the legitimate claims of all parties while disappointing everyone somewhat.
An algorithm cannot do this. An algorithm can only optimise for a metric. And whatever metric you choose will necessarily be incomplete. It will measure some important things and miss others. By outsourcing judgment to the algorithm, you are not improving judgment. You are replacing judgment with optimisation. And the difference is that optimisation is mindless while judgment involves understanding.
The final irony the Prat identifies is that automation of judgment does not actually reduce the burden on those deploying it. It shifts the burden. Instead of making difficult decisions, organisations now have to make the difficult decision of whether to trust the algorithm. They have to decide what metrics to optimise for. They have to decide whether the algorithm's output is acceptable. They have shifted the labour from deciding to auditing, from taking responsibility to managing systems.
This is, perhaps, an improvement in some contexts. But it is not the improvement that was promised. The promise was that algorithms would remove the burden of judgment. The reality is that algorithms shift the burden, distribute it, diffuse it, make it harder to locate. But the burden of getting decisions right has not disappeared. It has merely been concealed.
What the Prat's reading ultimately suggests is that we should be suspicious of any technology that promises to remove the burden of judgment. Judgment is not a burden to be removed. Judgment is the core of what it means to be responsible. The goal should not be to automate judgment but to improve it — to develop better ways of making difficult decisions, to make those decision-makers more accountable, to build in more oversight and more opportunity for appeal. The goal should be to make human judgment better, not to replace it with algorithmic approximation.
Algorithmic decision-making has been deployed in criminal justice (COMPAS recidivism prediction), credit (automated loan denial systems), hiring (résumé screening algorithms), and education (automated student placement). Studies have documented that these algorithms frequently reproduce historical biases present in training data while creating the appearance of objectivity. The responsibility structure of algorithmic decision-making is complex — programmers do not control how algorithms are used, organisations deploying algorithms can claim they are "following the algorithm," and affected individuals have limited ability to challenge algorithmic decisions. The European Union's AI Act attempts to impose explainability and accountability requirements on algorithmic decision-making, but implementation remains contested.
Auf Wiedersehen, amigo!