What makes something a dilemma
A moral dilemma is not just a hard decision. It is a situation where every available option gives up something that matters, so no choice is free of cost. People face these constantly. What is new with AI is that someone must decide in advance how a machine will act in such situations, and write that decision into software or into the training of a machine learning model.
Thought experiments are a useful way to see this. They strip away detail so that the underlying trade-off is visible, and they are a standard exercise in critical thinking. They are not predictions of what will happen, and real cases are usually messier. Treat the scenarios below as tools for thinking, not as descriptions of any particular system.
The trolley problem and its limits
The best-known example is the trolley problem: a runaway vehicle will hit five people unless it is diverted onto a track where it will hit one. Philosophers have used it for decades to separate two intuitions. One says to minimize total harm. The other says there is a moral difference between causing harm and allowing it.
It is often applied to machines, especially vehicles. But the scenario is deliberately artificial. In real situations the outcome is rarely certain, the options are not clean, and a system usually has to act on imperfect information. A more realistic question is how a system should behave under uncertainty, not which person it should choose. The specifics for road vehicles are covered in a separate piece on autonomous vehicle ethics.
The screening scenario: false alarms versus missed cases
Imagine a hypothetical tool that screens people for a rare problem, such as fraud or a medical condition. It can be tuned in two directions. Set it to be cautious and it flags many people who are actually fine. Set it to be lenient and it misses some who are not.
Neither setting is neutral. Flagging too often burdens, embarrasses or harms people who did nothing wrong. Flagging too rarely lets real problems through. Where to set the threshold is not a technical question with a single correct answer. It depends on how bad each kind of error is, and who bears the cost. This is a place where reading statistics carefully matters, since a tool that sounds highly accurate can still produce many false alarms when the thing it is looking for is rare.
The fairness scenario: equal treatment or equal results
Suppose a system helps decide who gets an interview. Two intuitions about fairness conflict. One says everyone should be judged by the same rules. Another says outcomes should be similar across groups, or at least that rules should not lock in past disadvantage. Researchers have shown that several reasonable mathematical definitions of fairness cannot all be satisfied at once, except in special cases. So designers must choose, and choosing is a value judgment as much as an engineering one.
That is part of why concerns about hiring bias are hard to resolve by technical fixes alone. Someone has to decide which kind of fairness to prioritize and defend that choice.
The privacy scenario: safety versus being watched
Consider a hypothetical system that scans activity in a public place to detect emergencies sooner. It might help people get aid faster. It also means everyone in the area is being monitored, including people who have done nothing unusual. Some people accept that for the safety benefit; others regard constant observation as a harm in itself, even if no one misuses the data. The tension between privacy versus security is old, and AI mainly changes the scale and speed at which watching can happen.
The autonomy scenario: helping versus deciding
Another dilemma concerns who is in charge. A system that recommends is different from a system that decides. If a tool, like the recommendation engines used in online shopping, nudges people toward options it predicts they will like, it may be helpful or it may quietly narrow their choices. If a tool makes a decision automatically because it is faster, people may lose the chance to explain their circumstances. Convenience and human agency do not always point the same way.
Why machines make this sharper
People have always faced these dilemmas, so what is different about AI? A few things stand out:
- Decisions are made at scale, so a small bias or error is repeated many times.
- The choices are made in advance, which forces explicit trade-offs that humans often handle by instinct in the moment.
- It can be unclear who is responsible, because the decision comes from a mix of data, design and deployment.
- The reasoning may be hard to inspect, so people cannot easily tell whether a trade-off was handled sensibly.
Ways people try to handle dilemmas
There is no formula, but a few practices help. One is to name the trade-off openly instead of hiding it inside a default setting. Another is to involve the people affected, since those who bear the cost may weigh it differently from designers. A third is to keep a person able to review and override decisions in high-stakes cases. A fourth is to test how a system behaves in unusual situations, not only typical ones, and to monitor it after release.
Philosophical frameworks can inform these choices. One focuses on outcomes, another on duties and rights, another on character and fairness of process. They often disagree, which is partly why dilemmas stay dilemmas.
What to take from the scenarios
The value of these thought experiments is not in finding the “right” answer. It is in noticing that when a machine acts, someone has already decided how to trade one good against another. Seeing that decision clearly is the first step toward asking whether it was made well, and by whom. If the answer to “who decided?” is “nobody in particular,” that is usually a sign that more care is needed.