Ethics of Artificial Intelligence

What it asks

A system trained on data recommends, ranks, approves or refuses. It causes real effects on real people. Nobody chose those effects individually, the people who built it cannot say exactly why it behaves as it does, and the institution deploying it did not write it.

That is a responsibility gap, and it is the field's characteristic problem. The system is not an agent, so it cannot be blamed. The engineers built a general capability, not the outcome. The deployer followed the documentation. The harm is real and nobody is obviously culpable.

Most of the field's other questions - bias, opacity, autonomy, consent - are versions of this one.

The recurring problems

Bias. A system trained on records of past decisions reproduces the pattern of those decisions. That is not a malfunction, it is the system working. Making it fair requires deciding what fairness is, and the leading formal definitions have been proved mutually incompatible - so it is a choice, not an optimisation.

Opacity. A decision that cannot be explained cannot be contested. Explanations that are produced after the fact may describe the reasoning or may be plausible stories, and telling those apart is unsolved.

The responsibility gap. If no person is culpable, either the harm is nobody's fault or responsibility must be assigned on grounds other than fault. Both options are uncomfortable and one has to be taken.

Moral status. Whether any artificial system could deserve consideration in its own right. It is not urgent now, and the reason to keep it in view is that the question will be asked long before anyone can answer it.

How it developed

Speculation about machine morality is old, and for decades it was the whole field: thought experiments about robots, largely disconnected from any working system.

The field became practical in the 2010s, when systems that learn from data were deployed at scale in hiring, lending, policing and content ranking. The problems that emerged were not the ones the speculative literature had prepared for - not machines with goals of their own, but ordinary statistical systems reproducing existing patterns of disadvantage at speed and with an appearance of objectivity.

Work on algorithmic fairness then produced the field's most solid result: the leading formal criteria of fairness cannot all be satisfied at once, so any deployed system embodies a choice among them.

More recently the debate has divided between those focused on harms occurring now and those focused on risks from much more capable future systems, and that split is as much about how to allocate attention as about what is true.

The positions within it

Consequentialist, deontological and virtue-based approaches all appear, as in any applied field. Machine ethics asks whether systems can be built to reason morally; value alignment asks how to specify what a capable system should pursue. Abolitionist positions hold that some applications should not exist rather than be improved. All are in the Schools section.

Where it stands now

The field has institutional weight now - regulation, standards, procurement - and the risk that comes with it, which is being used to legitimate what it was meant to examine.

The most useful current work is unglamorous: on what an explanation of an automated decision has to do to be worth anything, on who bears the cost of an error, and on the labour and environmental cost of the systems, which the ethical debate ignored for a long time while discussing hypothetical futures.