What happens when the machine becomes smarter than the human? Can super-intelligent artificial intelligence be controlled by us at all? Researchers from the Max Planck Society have dealt with this question in a new study and come up with a clear answer: With the current state of computer science, it would not be possible to monitor super-intelligent AI using an algorithm.
Already today there are AIs that perform tasks independently without the programmers understanding exactly how they learned it. Manuel Cebrian, co-author of the study, which was published in the “Journal of Artificial Intelligence Research”, and his team followed two approaches to how humans could control super-intelligent AI.
One approach would be to limit the AI’s resources by separating its access to the Internet and other technical devices. However, this also slows down the progress of AI – and its benefits are less.
The second approach would be to motivate the AI to only pursue goals in the interest of humanity and to teach it ethical rules. However, this control has its limits, as the researchers found in the study.
Control by algorithm not possible
They devised a theoretical algorithm that should ensure that AI does not harm humanity. In the simulation, the algorithm should stop the behavior of the AI if it considers it to be harmful. But the analyzes showed that according to the current scientific status, such an algorithm cannot be programmed. Complete control of the super-intelligent AI would not be possible as a result.
The algorithm would therefore bring its own processes to a standstill – and become unusable. The computer scientists can then no longer recognize whether the algorithm is still analyzing the AI or not doing anything, explains Iyad Rahwan, director of the Human and Machine Research Department. In addition, it is currently not possible to calculate whether a machine is more intelligent than a person. According to the researchers, this does not allow complete control of the AI by a human-designed algorithm.