How easy it is to control people

Researcher: Superintelligent AI cannot be controlled by humans in principle

In principle, super-intelligent artificial intelligence could not be controlled by humans. At least this is the conclusion reached by a group of scientists who have carried out theoretical calculations on the problem. In doing so, they went through two possible scenarios, they now explain. On the one hand, you could specifically withhold resources from super-intelligent AI, but then it would no longer be super-intelligent. On the other hand, you could program ethical rules directly into it, but that also has fundamental limits.

Science fiction closer than expected

As the researchers around Manuel Cebrian and Iyad Rahwan from the Max Planck Institute for Human Development explain, their analysis assumes that someone is programming an AI "whose intelligence would be superior to humans and which could learn everything independently". Connected to the Internet, it could not only access all human data, but also control machines. Based on their calculations, it would be fundamentally impossible to control such an AI so that it does not harm humanity. That sounds like science fiction, says Cebrian, but the first steps on the way have already been taken.

One way to keep such super-intelligent AI under control is to simply withhold resources from it right from the start, the researchers write. For example, you could deny it access to the Internet and operate it separately from machines in some other way. Without such contact with the outside world, however, it would be significantly less capable: "It would not solve the great problems of mankind in this way." Even minimal communication options could not guarantee complete security, but without such communication channels the whole purpose of such a machine would be reduced to absurdity.

Does the algorithm control AI?

An alternative would be to program an algorithm that monitors the super-intelligent AI and prevents harm to humanity. To do this, however, he would have to simulate the behavior of the AI ​​in order to intervene in good time if necessary. According to the current state of computer science, however, such an algorithm cannot be programmed in principle. In addition, we may not even be able to recognize if an AI is super-intelligent, the researchers write. They published their results in the Journal of Artificial Intelligence Research.

(mho)

Read comments (469) Go to homepage

Newsletter

Don't miss any news! Every morning the fresh news overview from heise online

E-mail address

You can find detailed information on the dispatch procedure and your cancellation options in our data protection declaration.

Ad ad