What “experiment” means here
The word experiment in the title can mislead, so to be clear: nothing was run, tested or measured for this article. There is no data, no survey and no simulation. It is a thought exercise, a way of walking through hard questions to see why they are hard. The scenarios below are hypothetical, and where people disagree about the right answer, the disagreement is real and unresolved.
Can a machine make a moral decision?
A machine does not have values the way a person does. It follows instructions, or it applies patterns learned from data. When a system makes a choice that has moral weight, such as approving a loan or deciding how to act in an emergency, the values involved come from people: those who designed it, chose its goals, selected its data and decided where to use it.
So the question is usually not “what would the machine decide?” but “what did the people building it decide in advance, and did they think about it carefully?” Moral choices made by machines are really moral choices made earlier, by humans, and embedded in code and data.
The trolley problem, briefly
The best-known puzzle is the trolley problem. A runaway trolley is heading toward five people on a track. You can pull a lever to divert it to a side track where it will hit one person. Many people say pulling the lever is acceptable. A variation asks whether you would push a large person off a bridge to stop the trolley, saving the five. Most people feel much more reluctant, even though the numbers are the same.
The puzzle was designed by philosophers to probe how we reason about harm, intention and responsibility. It reveals a split between thinking about outcomes (fewer deaths is better) and thinking about duties or rights (it is wrong to use a person as a means).
Why the trolley problem is a poor guide to real systems
It has become a popular way to talk about, for example, self-driving cars. But there are good reasons to treat it carefully.
- Real situations are uncertain. A vehicle rarely knows for sure what will happen next. It deals with probabilities, not clean choices between known outcomes.
- The best action is usually to avoid the dilemma. Engineers focus on braking sooner, driving more cautiously and detecting hazards early, so that these forced choices become rare.
- Different societies hold different intuitions. What people consider acceptable varies, and there is no universal consensus to program in.
- Responsibility is spread out. If something goes wrong, blame might involve the manufacturer, the owner, the regulator and others, which relates to automated accountability.
The puzzle is useful for sharpening thinking, but most ethical trouble in AI looks less dramatic.
Fairness questions: the everyday dilemmas
Here is a more realistic hypothetical. A bank uses a model to help decide who gets a loan. It is trained on past decisions. Several questions arise.
What counts as fair? One idea is that people with similar finances should get similar outcomes. Another is that approval rates should be similar across groups. A third is that the model’s mistakes should be equally likely for every group. These can conflict with each other, and in many situations it is mathematically impossible to satisfy all of them at once. Choosing between them is a value judgment, not a technical one.
What if past data reflects past unfairness? A model trained on history may learn to repeat it. This is a core source of algorithmic bias.
Should the system explain itself? A person turned down for a loan may reasonably want to know why. If the model is too complex to explain, that raises the stakes of explainable AI.
Who checks the result? In high-stakes decisions, many people argue that human oversight should be real, not a rubber stamp.
What about privacy? The data used to make decisions may be sensitive, which touches on data privacy.
Another hypothetical: the triage assistant
Imagine a system suggesting which of several patients should be seen first in a busy clinic. It might weigh urgency, age and chances of benefit. Reasonable people disagree on whether age should count at all, how to value a small chance of a big benefit against a large chance of a small one, and whether a machine should have any say. This is a general illustration, not a description of any real system, and in practice such decisions belong with trained professionals, with any tool serving as an aid.
Approaches people take
There is no agreed method, but a few broad ideas appear often.
- Rules and constraints. Define things the system must never do.
- Weighing outcomes. Try to choose actions that produce the best overall results.
- Process and rights. Focus on whether people were treated with respect and due process.
- Participation. Involve the people affected in setting the rules.
- Auditing and testing. Check how a system behaves across different groups and situations, before and after launch.
Each approach has strengths and blind spots, and real governance usually mixes several. Governments are also experimenting with AI regulation, and the wider field of AI safety studies how to keep systems aligned with human goals.
What this exercise suggests
Three modest takeaways follow. First, ethical questions about AI are mostly questions about human choices, such as what to build, for whom, with what data and under what checks. Second, dramatic dilemmas get attention, but quieter issues like fairness, transparency and accountability affect more people. Third, there is no setting that makes a system morally correct. The best available answer is careful design, honest testing, clear responsibility and the willingness to change course.
These are general observations, not findings, and thoughtful people will weigh the questions differently.