Creation of autonomously acting, learning artifacts has reached a point where humans cannot any more be justly held responsible for the actions of certain types of machines. Such machines learn during operation, thus continuously changing their original behaviour in uncontrollable (by the initial manufacturer) ways. They act without effective supervision and have an epistemic advantage over humans, in that their extended sensory apparatus, their superior processing speed and perfect memory render it impossible for humans to supervise the machine’s decisions in real-time. We survey the techniques of artificial intelligence engineering, showing that there has been a shift in the role of the programmer of such machines from a coder (who has complete control over the program in the machine) to a mere creator of software organisms which evolve and develop by themselves. We then discuss the problem of responsibility ascription to such machines, trying to avoid the metaphysical pitfalls of the mind-body problem. We propose five criteria for purely legal responsibility, which are in accordance both with the findings of contemporary analytic philosophy and with legal practise. We suggest that Stahl’s (2006) concept of “quasi-responsibility” might also be a way to handle the responsibility gap.
Since the dawn of civilization, man has lived together with artifacts: tools and machines he himself has called into existence. These artifacts he has used to extend the range and the quality of his senses, to increase or replace the power of his muscles, to store and transmit information to others, his contemporaries or those yet to be born. In all these cases, he himself had been the controlling force behind the artifacts’ actions. He had been the one to wield the hammer, to handle the knife, to look through the microscope, to drive a car, to flip a switch to turn the radio on or off. Responsibility ascription for whatever the machines “did” was straightforward, because the machines could not act by themselves. It was not the machine which acted, it was the controlling human. This not only applied to the simple tools, like hammers and knives, but also to cars and airplanes, remotely controlled planetary exploration vehicles and, until recently, computers.
Any useful, traditional artifact can be seen as a finite state machine: its manufacturer can describe its range of expected actions as a set of transformations that occur as a reaction of the artifact to changes in its environment (“inputs”). The complete set of expected transformations is what comprises the operating manual of the machine. By documenting the reactions of the machine to various valid input patterns, the manufacturer renders the reader of the operating manual capable of effectively controlling the device. This transfer of control is usually seen as the legal and moral basis of the transfer of responsibility for the results of the machine’s operation from the manufacturer to the operator (Fischer & Ravizza, 1998). If the operation of a machine causes damage, we will ascribe the responsibility for it according to who was in control of the machine at that point. If the machine operated correctly and predictably (that is, as documented in the operating manual), then we will deem its operator responsible. But if the operator can show that the machine underwent a significant transformation in its state which was not documented in the operating manual (e.g. by exploding, or failing to stop when brakes were applied) then we would not hold the operator responsible any longer, and precisely for the reason that he did not have sufficient control over the device’s behaviour to be able to assume full responsibility for the consequences of its operation.
With the advent of learning, autonomously acting machines, all this has changed more radically than it appears at first sight. Learning automata, as we will see, are not just another kind of machine, just another step in the evolution of artifacts from the spear to the automobile. Insofar as responsibility ascription is concerned, learning automata can be shown to be machines sui generis, in that the set of expected transformations they may undergo during operation cannot be determined in advance, which translates to the statement that the human operator cannot in principle have sufficient control over the machine to be rightly held responsible for the consequences of its operation.
Learning automata cause a paradigm shift in the creation, operation and evaluation of artifacts. In the progress of programming techniques from classic, imperative programming, to declarative languages, artificial neural networks, genetic algorithms and autonomous agent architectures, the manufacturer/programmer step by step gives up control over the machine’s future behaviour, until he finds her role reduced to that of a creator of an autonomous organism rather than the powerful, controlling coder that she still is in popular imagination and (all too often) in unqualified moral debate.
Key Terms in this Chapter
Imperative Programming: A programming paradigm where the programmer describes the machine’s actions step by step, thus keeping full control over the machine’s behaviour when executing the program.
Genetic Programming: A programming paradigm where the program “evolves” as a string of symbols out of other strings of symbols. The “evolution” process mimics the mechanics of biological evolution, including operations like genetic cross-over, mutations, and selection of the “fittest” program variants.
Epistemic Advantage: In general, the fact that an agent has by design better access to information about the world than another agent. In particular, the fact that some machines are in the privileged position to access data about the environment that humans cannot access (for example due to a lack of suitable sensor equipment, e.g. for gamma radiation or ultraviolet light); or that they are able to process information at a speed which transcends the speed of human thought, thus enabling them to handle situations in real-time, which humans cannot handle without machine aid (e.g. controlling a nuclear power plant, a low-flying fighter airplane, or a subtle orbital manoeuvre in space.)
Autonomous Agents: Programs or programmed devices which act autonomously, without supervision of a human, often in a remote location (e.g. on a remote server or another planet). Since such agents are per definition required to operate without supervision, responsibility attribution for their actions to a human is especially difficult.
Artificial Neural Network: A networked structure, modelled after a biological neural network, and implemented in software on a computer. Artificial neural networks enable computers to handle imperfect (noisy) data sets, which is essential for robust performance in advanced recognition and classification tasks (handwriting recognition, weather prediction, control of complex movements in robotic bodies).
Declarative Programming: A programming paradigm where the programmer does not specify the machine’s behaviour in detail. Instead, she describes the problem to be solved in a kind of predicate logic calculus, leaving the details of the inference process to the machine.
Learning Machine: A machine which modifies its behaviour after deployment, through adaptation to the environment in which it operates. Since its final behaviour at any moment depends not only on the initial programming, but also on the environment’s inputs, it is in principle not predictable in advance.