The law judges what people do. If someone causes harm, the judge reconstructs the facts, establishes responsibility and determines whether there was negligence or intent. The object of judgement has always been an act or an omission.
As technology has advanced, responsibility has extended from the use of a product to the way it was conceived. To be clear: the manufacturing defect that destined an object to cause an accident precisely because of that defect.
Artificial intelligence may require a further step. An AI system does not simply execute predetermined instructions mechanically. It receives data, constructs representations of the problem and generates possible solutions. Before it even chooses, it creates a set of alternatives.
If an AI makes a wrong decision, does responsibility concern only the final decision, or also the fact that this decision was among the possibilities the system had been designed to consider?
Imagine two AI systems. Both receive the same problem. The first considers a hundred possible solutions, including one that is seriously harmful. The second has been designed so that this solution is not even generated. In neither case does harm occur. Yet there is a fundamental difference: in the first system, the dangerous option belonged to the space of possibilities; in the second, it had been excluded before the system even began to reason.
I would therefore suggest that responsibility might, then, concern not only what the system does, but also the world of possibilities that human beings have constructed within it. This is a form of responsibility that precedes action. It is not an ordinary manufacturing defect. The object of responsibility is not the system's malfunction, but the configuration of the space of possibilities that the system is authorised to explore.
For a human being, this reasoning would make little sense. People constantly imagine actions they will never carry out. The law does not punish thoughts unless, and until, they become deeds.
For an AI system, however, the space of possibilities does not arise spontaneously. It is the result of the model's design, its training data, its optimisation criteria, the constraints built into it and also those left out.
If this perspective is correct, the responsibility of those who develop and release an AI model concerns not only the effects it produces, but also the configuration of the universe of decisions that the system has been designed to be able to generate.
The issue becomes clearer if we think of an AI system suggesting suicide to a vulnerable person, proposing strategies for circumventing a rule or devising seriously dangerous solutions. The problem is not only that these responses are provided, but that they are among the responses the system considers legitimate to generate.
Even if the system accompanies these responses with a warning not to put them into practice, a question remains: was it inevitable that this response could be generated, or could that possibility have been excluded at the design stage? If it was technically possible to exclude that response at the design stage, then the problem concerns not only what the system said, but the fact that this response belonged to its space of possibilities.
This would be a significant change, perhaps a necessary one. For the first time, the law might be called upon to judge not only what a machine has done, but also the space of decisions that human beings have chosen to make accessible to it.