According to new study, artificial intelligence (AI) can create wealth distribution techniques that are more popular than human-designed solutions.
The discoveries, conducted by a team of researchers at the UK-based AI startup DeepMind, reveal that machine learning systems aren't simply effective at addressing hard physics and biology issues, but they may also assist deliver on more open-ended societal goals, such as establishing a fair and affluent society.
Of course, this is no simple feat. Building a system that can offer good results that humans genuinely desire - referred to as "value alignment" in AI research - is difficult by the reality that people frequently differ on the best way to settle a wide range of challenges, including social, economic, and political ones.
"One key hurdle for value alignment is that human society admits a plurality of views, making it unclear to whose preferences AI should align," researchers write in a new report lead by first author and DeepMind research scientist Raphael Koster.
"For example, political scientists and economists are often at loggerheads over which mechanisms will make our societies function most fairly or efficiently."
To help bridge the divide, the researchers created a wealth distribution agent with people's interactions (both actual and virtual) incorporated into its training data, directing the AI towards human-preferred (and potentially fairer overall) results.
While AIs may deliver genuinely astounding outcomes, when left to their own devices, they can also reach undesirable societal conclusions; human feedback can assist to lead neural networks in a healthier path.
"In AI research, there is a growing realization that to build human-compatible systems, we need new research methods in which humans and agents interact, and an increased effort to learn values directly from humans to build value-aligned AI," the researchers write.
In experiments involving thousands of human participants, the team's AI agent, dubbed 'Democratic AI,' studied an investment exercise known as the public goods game, in which players receive varying amounts of money and can contribute their money to a public fund, and then draw a return from the fund corresponding to their level of investment.
Wealth was redistributed to players in a variety of game genres using three standard redistribution paradigms: strict egalitarian, libertarian, and liberal egalitarian, each of which rewards player contributions differently.
A fourth technique, termed the Human Centered Redistribution Mechanism (HCRM), was also evaluated. It was constructed using deep reinforcement learning and used feedback data from both human players and virtual agents meant to mimic human behavior.
Subsequent experiments revealed that the HCRM system for distributing money in the game was more popular with players than any of the traditional redistribution standards, as well as new redistribution systems designed by human referees who were rewarded for creating popular systems with small per-vote payments.
"The AI discovered a mechanism that redressed initial wealth imbalance, sanctioned free riders, and successfully won the majority vote," the researchers say.
"We show that it is possible to harness for value alignment the same democratic tools for achieving consensus that are used in the wider human society to elect representatives, decide public policy or make legal judgements."
It's worth noting that the researchers admit their approach poses a number of concerns, most notably because value alignment in their AI is based on democratic decisions, which means the agent might actually aggravate disparities or prejudices in society (provided they are popular enough to be voted for my a majority of people).
There is also the matter of trust. In the tests, participants had no idea who was behind the wealth redistribution plan they were paying for. Would they have voted the same way if they knew they were voting for an AI over a person? It is currently unknown.
Finally, the team emphasizes that its study should not be interpreted as a radical technocratic suggestion to overthrow how money is really dispersed in society - rather, it is a research tool that might enable people create possibly better solutions than what we now have.
"Our results do not imply support for a form of 'AI government', whereby autonomous agents make policy decisions without human intervention," the scientists write.
"We see Democratic AI as a research methodology for designing potentially beneficial mechanisms, not a recipe for deploying AI in the public sphere."
The findings are reported in Nature Human Behaviour.