DeepMind researchers have trained an AI system to find a popular policy for distributing public funds in an online game – but they also warn of “AI government”
July 4, 2022
A ‘democratic’ AI system has learned how to develop the most popular policies for redistributing public money among people who play an online game.
“Many of the problems people face are not just technological, but require that we coordinate in society and in our economies for the greater good,” says Raphael Koster at UK-based AI company DeepMind. “To help, AI needs to learn directly about human values.”
The DeepMind team trained its artificial intelligence to learn from more than 4000 people and from computer simulations in an online economic game for four players. In the game, players start with different amounts of money and must decide how much they want to contribute to grow a pool of public funds, eventually receiving a share of the pot in return. Players also voted on their preferred policy for handing out public money.
The policies the AI developed after this training generally sought to narrow the wealth disparities between players by redistributing public money based on how much of their starting pot each player contributed. It also discouraged free-riders by giving almost nothing back to players unless they contributed about half of their starting money.
This AI-designed policy won more votes from human players than either an “egalitarian” approach to distributing funds equally regardless of how much each person contributed, or a “libertarian” approach to distributing funds according to the share of the contribution of each person the public pot.
“One thing we found surprising was that the AI learned a policy that reflects a mixture of views from across the political spectrum,” he says. Christopher Summerfield at Deep Mind.
When there was the greatest disparity between players at the outset, a “liberal egalitarian” policy — which redistributed money according to the share of starting money each player contributed, but didn’t discourage free-riders — proved just as popular as the AI proposal. , by obtaining more than 50 percent of the vote in a head-to-head competition.
The Deep Mind researchers warn that their work is not a recipe for ‘AI government’. They say they have no intention of building AI-powered policy-making tools.
You can too, because the AI proposal isn’t necessarily unique compared to what some people have already suggested, says Annette Zimmermann at the University of York, UK. Zimmermann also warned against focusing on a narrow idea of democracy as a “preferential gratification” system for finding the most popular policies.
“Democracy isn’t just about winning, implementing whatever policies you want, it’s about creating processes where citizens can meet and consult as equals,” said Zimmermann.
The DeepMind researchers express concern about an AI-powered “majority tyranny” in which the needs of people in minority groups are overlooked. But that’s not a big concern among political scientists, says Mathias Risse at Harvard University. He says modern democracies face a bigger problem of ‘the many’ being disenfranchised by the small minority of the… economic eliteand step out of the political process altogether.
Still, Risse says the DeepMind study is “fascinating” in the way it has delivered a version of liberal egalitarianism policies. “Because I’m in the liberal-egalitarian camp anyway, I think that’s a pretty satisfying result,” he says.
Reference magazine: Nature Human behavior† DOI: 10.1038/s41562-022-01383-x
More on these topics: