Artificial intelligence algorithms are sometimes black boxes with inaccessible rules. To understand the behavior of the machines that have them, we must create a new scientific discipline, as we have created the study of animal behavior. This is the point of view of Jean-François Bonnefon who, along with 22 other scientists, has just signed a tribune in the journal Nature.
Our social, cultural, economic and political interactions are becoming increasingly important to a new type of actors: machines with artificial intelligence. These machines filter the information that reaches us, guide us in the search for a spouse, and converse with our children. They exchange securities on the financial markets, they advise judges and police. Soon, they will drive our cars and make war on us. If we want to keep these machines under control, get the greatest benefits and minimize potential damage, we need to understand their behavior.
Understanding the behavior of smart machines is a broader goal than understanding their programming. Sometimes the programming of a machine is not accessible, for example, when its code is an industrial secret. In this case, it is necessary to understand a machine from the outside, observing its actions and measuring its consequences. Other times, it is not possible to completely predict the behavior of a machine from its code, because this behavior will change in a complex way when the machine will adapt to its environment, by a process of learning guided but ultimately opaque. In this case, it is necessary to observe this behavior continuously and to simulate the potential evolutions. Finally, even when one can predict the behavior of a machine from its code, it is difficult to predict how the actions of the machine will change the behavior of humans (which are not programmable), and how human actions are going to change the behavior of the machine. In this case, it is necessary to conduct experiments to anticipate the cultural coevolution of humans and machines.
New science to observe machines:
InToeet all these challenges, we need to create a new scientific discipline, dedicated to the behavior of machines, as we have created the scientific discipline of animal behavior. The behavior of animals can not be understood solely based on genetics, organic chemistry or cerebral anatomy: we also need observational and experimental methods, studying the animal in its environment or the laboratory.
“It is difficult to predict how the actions of the machine will change the behavior of humans and how the actions of humans will in turn change the behavior of the machine.It is then necessary to conduct experiments to anticipate the cultural coevolution of humans and machines.”
In the same way, we can not understand the behavior of intelligent machines solely based on computer science or robotics: we also need behavior specialists trained in experimental methods in the fields of psychology, economics. In the same way, we can not understand the behavior of intelligent machines solely based on computer science or robotics: we also need specialists in the field of science. never created from scratch. The behavior of animals was studied by many scientists long before the study of animal behavior was formalized as a structured and independent discipline. In the same way, many scientists will recognize themselves in the discipline of the behavior of the machines, once this discipline will be structured and identified. But the most important thing is that they will recognize each other, much more easily than today. They are trained in experimental methods in the fields of psychology, economics, political science or anthropology.
This is nowadays to collapse the posters and to mutually identify and to organize their direction across the current disciplinary boundaries; we will allow public authorities and regulatory agencies to rely more easily on a body of science today dispersed and difficult to access; and we will allow citizens to guide and enlighten the world through the emergence of intelligent machines.
One can not predict 100% of the behavior of robots who constantly learn about their interactions with their environment. According to Jean-François Bonnefon, we need to create a science of the behavior of machines to observe them experimentally.
This is the motivation of a call to researchers, public decision-makers, companies that shape the intelligent machines, published in the journal Nature by23 European and American coauthors, computer scientists, sociologists, biologists, economists, engineers, political scientists, anthropologists and psychologists, researchers in public research organizations, academics, or employees of the giants of artificial intelligence that are Microsoft, Facebook or Google. We look at the big questions that could be the basis of the machine behavior champion, inspired by the questions that were based on the champion of animal behavior.
How is behavior shaped and evolved?
Among these big questions, there are the social and economic incentives that have waited for the expected behavior. For example, what is the metric that has been optimized, an algorithm for verifying information on social networks, and what are the unexpected psychosocial effects of this initial goal?
Another type of big question: by what mechanism has behavior been acquired and by what mechanisms has it changed? For example, on what type of data was a predictive policing algorithm quickly? If these data have been biased against a particular social group, is the algorithm likely to magnify this bias by its decisions and so be in a spiral of injustice?
“By bringing together what is today scattered (…), we will allow citizen citizens to remain more lucid in a world shaken by the emergence of intelligent machines.”
Knowing in what environment can we change or spread, and in what environment is it already gone, has it also responded to the big questions we examined? For example, an open archive of algorithms for autonomous cars may be-programmed-model-of-car-of-oneself-is-more-fast-of-new-model-of -model-to-day
All these questions must be presented at the scale of the isolated machine, the machine to interact with other machines, and at the scale of hybrid collectives. All are essential, but today they are ordered by order dispersed by the communities that are distinguished to surrender. Bringing these communities together under the banner of the new science of machine behavior will be a decisive step towards harmoniously turning the tide of artificial intelligence.
