An “democratic” AI system can learn how to create the most popular policy in redistributing public funds among online gamers.

Raphael Koster, an AI company based in the UK, says that many of the problems facing humanity are more than technological. “We need to coordinate in society as well as in our economies for greater good.” “AI must learn directly about human values in order to be able help,” says Raphael Koster, UK-based AI company DeepMind.

DeepMind’s artificial intelligence was trained by the DeepMind team to learn from over 4000 people and from computer simulations in an economic game online. The game requires players to start with different amounts and decide how much they will contribute to grow the public fund pool. In return, they receive a portion of the pot. The game also allowed players to vote on their preferred policies for disbursing public money.

After this training, the AI developed a policy that aimed to reduce wealth disparities among players. It distributed public money according to how much each player contributed to their starting funds. The AI also discouraged free-riders, giving almost nothing back to players unless they contributed at least half of their initial funds.

The AI-designed policy received more votes than any other “egalitarian” approach that distributes funds equally, regardless of how much each person contributed or a “libertarian approach that distributes funds according to what percentage each person contributes to the public pot.

DeepMind’s Christopher Summerfield says, “One thing that we found surprising was the AI learning a policy that reflects an assortment of views from across all the political spectrum.”

A “liberal egalitarian”, which redistributed money according the player’s contribution to starting funds, but did not discourage free-riders, was popular when there was highest inequality among players. It won more than 50% of the vote share in a head–to-head contest.

DeepMind researchers warn their work is not a blueprint for AI government. They claim they have no plans to create AI-powered tools for policy-making.

This may be true, as the AI proposal is not unique in comparison to what others have suggested, Annette Zimmermann, University of York, UK, states. Zimmermann warned that it was not a good idea to view democracy only as a system of preference satisfaction for choosing the most popular policies.

Zimmermann says that democracy is not about winning or getting the best policy implemented. It’s about creating processes where citizens can meet and discuss their issues as equals,” he said.

DeepMind researchers raise concerns about an AI-powered “tyranny by the majority” situation where the needs of minorities are ignored. Mathias Risse, a Harvard University political scientist, said that this is not a major concern. According to Risse, modern democracies are facing a greater problem in which “the many” become disenfranchised and drop out of the political process by a small number of the economic elite.

Risse still believes DeepMind’s research is “fascinating” because it produced a liberal egalitarianism version. He says, “Since i’m in liberal-egalitarian camp anyways, I find this a rather satisfactory result.”