Artificial intelligence will imprison humanity

Would a digital super intelligence be controllable?

Out of control: How realistic are science fiction scenarios of uncontrollable, human-damaging artificial intelligences? According to computer scientists, it might actually be impossible to have a self-learning, super-intelligent AI system fully monitor. Even a control by special protection algorithms could be ineffective for programming reasons.

Artificial intelligence is making rapid advances. Thanks to neural networks and adaptive algorithms, AI systems are already controlling autonomous vehicles, diagnosing diseases and deciphering the complex structure of proteins. Even bicycling, poker games and strategy games are already able to master adaptive machine brains today - sometimes even better than we humans. At the same time, their decisions are sometimes less objective than one would expect.

Super AI coming soon?

This raises the question of how far this development will go: Will there ever be artificial intelligence (AI) that is superior to humans? Connected to the Internet, such an AI would have access to all human data, it could replace all existing programs and control all machines connected to the Internet. Would such an AI system cure cancer, bring about world peace and prevent climate catastrophe? Or would it destroy humanity and take over the earth?

"A super-intelligent machine that controls the world sounds like science fiction," says co-author Manuel Cebrian from the Max Planck Institute for Human Development. “But even today there are machines that perform certain important tasks independently, without the people who programmed them fully understanding how they learned it. Therefore, the question arises for us whether this could at some point become uncontrollable and dangerous for humanity. "

This is already being debated in connection with autonomous weapons, which could kill people or start a war simply because of the speed at which they make decisions before people are able to intervene.

Lock up or control by rules?

Cebrian, first author Manuel Alfonseca from the Autonomous University of Madrid and her colleagues have theoretically worked out whether and how a super AI could be controlled. Accordingly, there are two basic approaches to controlling a digital superintelligence: On the one hand, its resources could be deliberately restricted by, for example, isolating them from the Internet and other devices so that they have no contact with the outside world. But this would also mean that the capabilities of the super-intelligent AI would be significantly lower. It would not solve the great problems of mankind in this way.

The other option would be to motivate the AI ​​from the outset to only pursue goals that are in the interests of humanity, for example by programming ethical rules into it. Well-known examples of such rules are Isaac Asimov's three robot laws. In short, they say that a robot must not harm people, that it should obey people, unless the order violates rule 1. And the robot is supposed to protect itself, unless it contradicts rules 1 and 2.

The problem, however, is that such rules are suitable for simple algorithms, but not for self-learning, super-intelligent AI. "Because such a super-intelligence can mobilize a wide variety of resources in order to achieve goals that are potentially incomprehensible to humans and certainly not controllable," explain the researchers.

Control algorithm stops itself

Alfonseca and his team use an example to illustrate where the problems lie in practice: They are designing a theoretical algorithm that is supposed to ensure that super-intelligent AI does not harm humanity under any circumstances. To do this, this algorithm first has to understand the behavior of the AI ​​and, in a kind of dry run, find out whether its actions would be harmful. If this is the case, a second subroutine follows, which stops the AI.

In practice, however, this principle reaches its limits: “If you break the problem down to simple basic rules from theoretical computer science, it turns out that an algorithm that would order an AI not to destroy the world accidentally brings its own processes to a standstill could bring ”, explains Cebrian's colleague Iyad Rahwan. "That makes this algorithm practically unusable."

"Basic Mathematical Limitations"

On the basis of their calculations, the scientists come to the conclusion that it is not possible to program an algorithm that recognizes whether an AI would damage the world or not - and still work. "There are basic math limitations to using one AI system to keep a second on a zero-risk course," they explain.

In addition, we probably wouldn't even know if a machine was super-intelligent. Because whether the intelligence of such an AI is superior to humans cannot be calculated based on current knowledge. (Journal of Artificial Intelligence Research, 2021; doi: 10.1613 / jair.1.12202)

Source: Max Planck Institute for Human Development, Berlin

January 22, 2021

- Nadja Podbregar