The Ethics of Artificial Minds

A recommendation system is a moral object. It decides what millions of people see, it optimises for something someone chose, and its effects are unevenly distributed. The interesting thing is that we already have three well-developed frameworks for assessing it, and they do not agree, which is exactly what makes them useful.

Start with the utilitarian question: does the system increase total wellbeing? Sum the benefits and harms across everyone affected, and prefer the arrangement with the better total. This is the framework that most closely matches how such systems are actually evaluated, since aggregate metrics are what dashboards produce.

Its weakness is structural and shows up quickly. Aggregation can absorb concentrated harm. Suppose a change raises overall engagement, and the increase is concentrated in a small group showing signs of compulsion. The totals improve. A strict utilitarian calculation approves, and the people being harmed are averaged away by people being mildly entertained.

Now the Kantian question: does the system treat people as ends in themselves, or merely as means? Kant's claim is that rational beings set their own ends, and that this gives them a worth which is not exchangeable for anything. Using someone as a means is ordinary and fine: that is what employing a taxi driver is. The wrong is treating them as nothing but a means.

Applied to software this is sharper than it sounds. Terms written so that nobody will read them produce formal consent while being designed against real understanding, which bypasses precisely the capacity Kant says gives a person their standing. An interface optimised to defeat a user's own intentions about their time relates to them as an obstacle. The Kantian objection does not depend on outcomes at all, which is why it survives good engagement numbers.

Third, the Rawlsian question: would you accept this system without knowing which user you would be? Not the average user: the one it ranks lowest, flags most often, or serves worst. This is the original position applied to a product, and it is the most directly usable of the three, because it converts a fairness question into a concrete exercise a team can actually run in a meeting.

Watch the three come apart on one case and you have the whole value of knowing them. The engagement change above: utilitarianism may approve on the totals, Kant objects to the relation the design establishes with the vulnerable group, and Rawls objects because you would not accept the system from that group's position. When frameworks disagree, they are telling you where the actual decision is, rather than dressing a decision you already made.

There is a fourth thread running under all of this, and it is older. A hiring model trained on a decade of past decisions predicts future success from past regularity. That is precisely the inference Hume showed cannot be justified: the training data is entirely historical and the claim is about tomorrow. Ordinarily this is a philosopher's problem. Here it is an operational one, because if the past hiring encoded a bias, the model does not merely inherit it. It launders it, presenting a contingent human pattern as a law discovered in data.

Which is the practical case for reading any of this. The frameworks are not decoration on a technical problem. They are the only vocabulary anyone has developed for the questions the technical work cannot answer, and those are the questions that turn out to matter.