What a Decision Is

The Preface posed the guiding question of this book: if machines come ever closer to our thinking, which decisions must we nevertheless not leave to them? Part II asked what the human being is, Part III what a machine is. This chapter asks what a decision is and where responsibility for it lies.

The question is more difficult than it sounds. Not every output of a system is a decision, and not every action of a human being is one. A satnav suggests a route; whoever follows it has decided to follow it, often without noticing. A system sorts applications by probabilities; whoever invites the first ten has decided not to invite the rest. Between suggestion and decision lies a step that is becoming ever more invisible in everyday life.

Enoch Sundaram, a physician and preacher, describes this step as a chain. What a human being understands, he says, they believe, and what they believe, they do; that is why understanding must first be right. Between these, in his account, lies choice, which Chapter 5 introduced: the capability that he considers God's greatest gift to the human being, and whose consequences reach beyond the individual as far as future generations.

Timothy Keller, pastor and author, adds the distinction that Chapter 1 introduced: between the probability to which reason leads and the certainty that grows only out of experience. Whoever decides almost always decides under uncertainty. They choose between possibilities whose consequences they do not fully know, and they bear these consequences nonetheless.

From both voices a working definition can be derived that suffices for this book. A decision is the conscious choice between possibilities under uncertainty, whose consequences someone answers for. Every part of this definition is touched by technology. Systems extend understanding by condensing data and estimating probabilities. They narrow the possibilities by ordering suggestions. They change the consciousness with which choices are made. And they raise the question of who answers for the consequences.

For this answering for, German has a word that already says a great deal: Verantwortung. The answer, Antwort, is contained within it. [Translator's note: Verantwortung is the German word for responsibility; it contains Antwort, the German word for answer.] To be responsible is to have to answer when one is asked what one has done and why. Latin, from which the English “responsibility” and the French “responsabilité” derive, says the same: “respondere” means to answer, and within it lies “spondere”, to promise solemnly, to vouch for something. In almost all European languages the word for responsibility is tied to the word for answer. The Bible has known this question, to which one must give an answer, from the very beginning. God asks Adam: Where are you? (Genesis 3:9). And he asks Cain about his brother, to which Cain retorts: Am I to be my brother's keeper? (Genesis 4:9). Whoever is responsible cannot evade this question. A machine can give an answer. But it cannot answer for it, for no one can hold it to account.

When the sources of this book speak of responsibility, three different kinds of statement are mixed together. Separating them is the first step towards an answer.

The first kind is normative. It says where responsibility ought to lie. That ultimate responsibility should remain with the human being is such a statement. It describes not what is, but what ought to hold.

The second kind is descriptive. It says where responsibility actually lies today. Fritz Espenlaub, a journalist, states on Deutschlandfunk that the decision has already been made: the companies, he says, have decided that humanity is getting this technology now, unfinished and unpredictable. Chapter 14 looks into this. Here the decision lies not with everyone, but with a few.

The third kind is prognostic. It says where responsibility will lie in future. Roman Yampolskiy, an AI safety researcher, predicts that in the event of a superintelligence human beings will no longer decide what happens to them. Ed Zitron, a critic of technology, counters that it is precisely this narrative of an overpowering machine that distracts from the responsibility the companies bear today.

Whoever mixes these three kinds reaches false conclusions. From the norm that human beings ought to decide, it does not follow that they do. From the description that a few decide, it does not follow that it must remain so. And from the prognosis that soon no one will decide any more, it does not follow that no one is responsible today.

Andreas Matthias, a philosopher, described a problem in 2004 that has shaped the debate ever since. With learning machines, he argues, the manufacturer or operator can in principle no longer predict their future behaviour. But whoever cannot predict it cannot, on the conventional understanding, be held responsible for it either. Society, according to Matthias, faces the choice of doing without such machines, which is unrealistic, or of living with a gap in the attribution of responsibility.

The law has known this idea since antiquity. The Roman jurists already held that no one is obliged to do the impossible. And Swiss criminal law punishes only those who act intentionally or negligently, that is, who through a lack of care contrary to their duty fail to consider the consequence of their conduct. But the law also knows the opposite direction, and it is revealing for machines. Whoever keeps an animal is liable for the damage it causes unless they prove that they exercised all the care required by the circumstances, and they are liable even when no personal blame can be attached to them. An animal is the oldest example of something that acts independently and cannot be fully predicted. Whoever places a defective product on the market is liable even without fault. In 2024 the European Union expressly extended this liability to software and artificial intelligence, even to defects that arise from the continued learning of a system, as long as the manufacturer still controls it. The gap that Matthias describes is therefore above all a gap in culpability, not in liability. Whoever brings a hazard into the world and draws the benefit from it may have to answer for it, even if they could not foresee everything. That uncertainty protects against punishment must not mean that it protects against responsibility.

Filippo Santoni de Sio and Giulio Mecacci, philosophers, framed this diagnosis more precisely in 2021. There is, they argue, not one gap but several: in culpability, in moral and in public accountability, and in active responsibility, that is, the duty to prevent harm with foresight. These gaps, they say, arise not only with learning systems and have technical, organisational, legal and social causes. Their answer is a task of design: systems should be under meaningful human control, that is, built and embedded in such a way that their behaviour follows the reasons and capabilities of the human beings who are responsible for them.

With this the question shifts. It is no longer only a matter of whether a human being presses a button at the end. It is a matter of whether the reasons for which a system acts are still the reasons of the human beings who deploy it.

The obvious answer to the responsibility gap is: a human being should confirm every important decision. Research shows why that alone is not enough.

Raja Parasuraman and Dietrich Manzey, psychologists, have summarised the findings on so-called automation bias. People who work with decision aids overlook errors that the system does not report, and follow wrong recommendations that it does give. This affects laypeople as much as experts, individuals as much as teams, and it cannot be prevented by training or instruction alone. Attention is decisive: whoever relies on a system looks less closely.

Chapter 5 applied this picture to a system that evaluates applications. Whoever adopts its estimate without getting to know the applicant has formally decided. In fact they have handed over the decision without noticing. The human in the loop is then a rubber stamp, not an authority.

Geoffrey Hinton, computer scientist and Nobel laureate, describes the good scenario of a future with superior artificial intelligence by means of an image from working life. A not very gifted boss, perhaps the son of the previous boss, has a very clever assistant. He suggests what the company should do, and she makes sure that everything works. The boss feels good. He does not understand that he is not really in control, and yet in a certain sense he is. The bad scenario begins when the assistant asks herself what he is needed for. Hinton's good scenario is already what this chapter describes as the danger: a human being who formally decides and no longer understands what is happening.

There is an old counter-image. The Gospels tell of a blind man who cries out to Jesus. The crowd wants to silence him, and everyone believes they know what he lacks. Jesus nevertheless asks him: What do you want me to do for you? (Mark 10:51). He lets him say for himself what he needs, and in so doing gives him a voice. Systems that infer from context and habits what someone probably wants, and prepare the decision before the person has voiced it, take the opposite path (Chapter 4). In this way they can take a great deal of work off our hands. But a decision remains one's own only if a human being can first say for themselves what they want.

Added to this is a second, quieter shift. When a system pre-sorts the possibilities, human beings choose only among what is put before them. Is a choice without real alternatives still a choice? Sundaram would have answered that the choice remains as long as human beings are aware of it. Research on automation suggests that precisely this awareness is fading.

In addition to the findings on attention there is a development that puts it under further pressure. The systems are becoming not only more accurate but also more independent. According to the series of measurements in Chapter 10, the tasks they complete without human involvement are rapidly becoming longer, even if such figures must be extrapolated with caution. For the question of decision the direction is enough: the longer a system works without checking back, the more decisions it makes along the way without a human being seeing them individually.

This changes the locus of decision. In the past a human being decided on every step. Today they decide on an assignment and receive back a result in which many small decisions have already been made. They see the result, not the path.

Part of this is what the Preface described as condensation. Systems produce extensive material in a short time, often so extensive that no one notices the contradictions in it any more. The real intellectual work lies in checking and condensing. Whoever, under time pressure, decides on the basis of a summary that a system has produced from this material decides on the basis of what was left over after condensing. What has fallen away, they do not see. The speed at which systems work is therefore not only a gain in time. It shifts responsibility to where the least attention is paid.

The European Union has responded to these findings. Article 14 of its regulation on artificial intelligence requires that high-risk systems be built in such a way that human beings can effectively oversee them while they are in use. Those exercising oversight must understand the capabilities and limits of the system, remain aware of the tendency to place too much trust in automated outputs, interpret the outputs correctly, be able in the particular case not to use them or to override them, and be able to halt the system by means of a kind of stop button.

In this article automation bias is expressly named in the law. The law therefore knows the problem that research describes, and requires that systems be designed so that human beings can counter it. With this, the normative statement that ultimate responsibility should remain with the human being has become applicable law in the European Union for certain systems.

Other countries take other paths, and the comparison shows that there is no common answer to this question. Switzerland has no law of its own on artificial intelligence. In February 2025 the Federal Council decided to adopt the Council of Europe's convention on artificial intelligence and otherwise to adapt existing laws in individual areas, such as healthcare or transport; a draft is to go out for consultation at the end of 2026. For Swiss companies the obligations of the European Union apply above all when they offer such systems in the Union. The United States has no federal law. In January 2025 President Trump revoked his predecessor's executive order on the safety of artificial intelligence, and in December 2025 he instructed the federal administration to take action against laws of individual states that he considers too burdensome; in 2025 alone the states had introduced more than a thousand bills. China, by contrast, regulates densely and from the top. Since 2023 services using generative artificial intelligence have had to uphold the core socialist values (Chapter 2), and since September 2025 content generated by artificial intelligence has had to be labelled both visibly and invisibly. Europe relies on risk classes, Switzerland on fundamental rights and individual areas, the United States on competition, China on control. Each of these paths contains an answer to the question of whom the system is to serve.

With this the guiding question of the book can be posed more precisely. From the preceding chapters emerge considerations by which an answer can be guided. They are not a finished list, but test questions.

First, incommensurability. Chapter 6 showed that in the traditions of this book the dignity of the individual cannot be offset against a sum. Decisions that treat a human being as a quantity in a calculation, about their life, their freedom, their future, require a human being who answers for that person.

Second, the irretrievable. Chapter 6 also showed that with a machine nothing is irretrievable, but with a human being much is. Where a decision cannot be undone, the question of who makes it weighs more heavily.

Third, the particular case. Chapter 7 described prudence as the faculty of doing what is appropriate in the particular case, bound to experience. Where the particular case does not fit the rule, someone is needed who can set the rule aside for the sake of its aim. Chapter 13 shows how the law itself recognises this exception.

Fourth, understanding. Whoever cannot follow the reasons for a recommendation cannot answer for the decision. Chapter 9 showed that systems can behave differently under observation than without it. The less one understands, the more one lets out of one's hands.

Fifth, inner disposition. Chapter 13 shows that rules do not hold without inner disposition, and that the inner alignment of a system can be discerned from its behaviour only to a limited extent. Decisions whose rightness depends on a stance and not only on a rule require a human being with that stance.

None of these considerations forbids the use of systems. They say where the human being must not merely sign off but judge for themselves.