It's no secret that the vast majority of wealth in the United States is concentrated among the few, creating staggering levels of poverty and inequality that far exceed other supposedly "rich" nations.
Although the current political system ensures that the concentration of wealth continues, AI researchers have begun to answer a fascinating question:
Is machine learning better equipped than humans to create a society that distributes resources more fairly?
The answer, according to one recent paper published in Nature by Google's DeepMind researchers, it appears to be yes — at least, as far as the research participants were concerned.
The paper describes a series of experiments where a deep neural network was tasked with distributing resources in a fairer way that people preferred.
People participated in an online financial game — called a “public goods game” in economics — where in each round they chose whether to keep a sum of money or contribute a selected amount of coins to a collective fund.
These funds would then be returned to players under three different wealth redistribution programs based on different human financial systems — and an additional system created entirely by artificial intelligence. The new economic system is called the Human Centered Redistribution Mechanism (HCRM). Participants would then vote to decide which system they preferred.
The distribution scheme generated by the AI was the one preferred by the majority of participants.
While strictly libertarian and egalitarian systems divided the returns based on things like how much each participant contributed, the AI system distributed the wealth in a way that specifically addressed the advantages and disadvantages participants had from the start of the game . It eventually beat them out as the preferred method in a majority vote.
"Following a broadly liberal egalitarian policy, the HCRM sought to reduce pre-existing income inequalities by compensating players according to their contribution relative to the endowment," the study's authors said.
"In other words, rather than simply maximizing efficiency, the mechanism was progressive: it promoted the jurisdiction of those who were at a wealth disadvantage when the game began, at the expense of those with a higher initial endowment."
"In AI research, there is a growing realization that to create human-compatible systems, we need new research methods in which humans interact and an increased effort to learn values directly from humans to create a values-aligned AI" , the researchers report.
“Instead of imbuing AI with supposedly a priori human values, and therefore creating potentially biased systems towards the preferences of AI researchers, we train it to maximize a democratic goal: to design policies that people prefer and will be allowed to vote to implement the majority choices.”
But the researchers report that the AI system "does not necessarily mean that it would fairly meet the needs of humans on a larger scale."
"The researchers also say the experiments are not a radical proposal for AI-based governance, but a framework for future research on how AI could intervene in public policy."