The DeepMind team has published a finding that AI is not only suitable for solving complex physical and biological problems, but also for creating a just and well-functioning society.
The problem is difficult because the machine is “value-modulated”, that is, it produces the result that people consider good. In the case of social issues, on the other hand, human opinions regarding social, economic and political issues are divided on well-known worldviews.
The main problem with aligning values is that human society recognizes different viewpoints, so it is not clear who AI should match with.
. said The nature of human behavior In the columns of Raphael Coster who leads his research.
Political scientists and economists often differ on the mechanisms that make society function more justly or more efficiently.
he added.
The problem was bridged by putting the machine into a game based on social redistribution, where it learned from human interactions between people – thus finding solutions that people find generally acceptable.
An important lesson in scientific work is that AI left alone leads to unacceptable results, so human participation is very important.
In AI research, recognition of the need to build human-compatible systems, a research method where humans and agents work together, and greater emphasis is placed on the values of learning AI directly from humans
Researchers wrote.
In the experiment, the Democratic Agent of Artificial Intelligence played an economic sports game with thousands of human participants. The essence of the game is that it forms a model for the operation of the common ownership: each player contributes to the common ownership based on his own decision, and then the common profit is distributed. In this game, there are participants who follow a cooperative strategy who contribute to the common, and there are “free runners” who participate in the common profit without investment.
There were various known methods of redistribution, strict egalitarianism, egalitarianism and liberalism, i.e. egalitarianism, libertarianism and the transition between the two.
The fourth method of redistribution was a human-centered redistribution mechanism invented by the machine. It was based on the strategies of human and robot players. During the experiments, it was revealed that human-centered redistribution was the most popular game mode, in contrast to the classic distribution logic described above and newer, popular-centered experimental methods.
The method of distributing the AI corrected income inequality and punished hikers, making it so popular that it won a majority of the vote.
The researchers expressed some reservations about the results. One is that AI operates according to a democratic logic, so depending on the votes, it can even drive the game toward increasing inequality and biases. The other was the question of the trust effect, because the players didn’t know they were voting for a solution to the distribution offered by the machine.
The result does not mean that we support some kind of state based on artificial intelligence. Democratic AI is a research method for designing potentially useful processes, not a recipe for using AI in the public sector
they mentioned.
(Cover Photo: Vernon Yuen/NurPhoto/Getty Images)
