AI Seems To Be Better At Distributing Wealth Than Human Beings Are, Research Hints

A new study proposes that artificial intelligence (AI) can devise approaches to wealth distribution that are more popular than systems developed by people.

The discoveries made by a group of researchers at UK-based AI firm DeepMind show that machine learning systems are not just proficient at fixing complicated physics and biology issues but might also assist supply on more open-ended social goals, such as the goal of realizing a fair, prosperous society.

Indeed, that is not a simple task. Constructing equipment that can give beneficial outcomes humans want– called “value alignment” in AI study– is complicated because individuals often differ on the most effective approach to solve all kinds of things, specifically social, economic, and political problems.

“One crucial hurdle for value alignment is that human society admits a plurality of sights, making it uncertain to whose preferences AI should align,” scientists describe in a new paper, led by initial author and DeepMind research scientist Raphael Koster.

“For example, political researchers and economists are commonly at loggerheads over which mechanisms will make our societies work most fairly or effectively.”

To help bridge the gap, the scientists developed an agent for wealth distribution that had people’s interactions (real and virtual) built into its training data– effectively directing the AI in the direction of human-preferred (and hypothetically fairer general) results.

Although AIs can create impressive results, they can also arrive at far-from-desirable social conclusions when left to their own devices; human feedback can assist in steering neural networks in a better way.

“In AI study, there is an expanding understanding that to construct human-compatible systems, we require new study methods in which humans and agents interrelate, and an increased effort to learn values directly from people to construct value-aligned AI,” the researchers write.

In experiments involving thousands of human individuals in total, the group’s AI agent– called ‘Democratic AI’– studied an investment exercise named the public goods game, in which players obtain varying amounts of money and might contribute their money to a public fund, and afterward draw a return from the fund corresponding to their level of investment.

In a series of distinct game styles, wealth was redistributed to gamers through three conventional redistribution paradigms– strict egalitarian, libertarian, and liberal egalitarian– each of which awards gamer investments differently.

A fourth technique was also evaluated, called the Human Centered Redistribution Mechanism (HCRM), created using deep reinforcement learning, using feedback information from both human gamers and virtual agents made to copy human behavior.

Subsequent experiments revealed that the HCRM system for paying out money in the game was more popular with gamers than any of the traditional redistribution standards and more popular than new redistribution systems designed by human referees who were incentivized to produce popular systems by getting small per-vote payments.

“The AI discovered a system that redressed initial wealth discrepancy, sanctioned free bikers, and effectively won the majority vote,” the scientists explain.

“We show that it is feasible to harness for value alignment the same democratic tools for achieving an agreement that is used in the larger human society to elect representatives, decide public policy or make legal judgments.”

It is worth noting that the researchers recognize their system raises several inquiries, mainly that value alignment in their AI revolves around democratic determinations, signifying the agent could currently exacerbate inequalities or biases in society (provided they are popular sufficient to be elected by a majority of people).

There is also the issue of confidence. In the experiments, gamers did not know the identity behind the wealth redistribution design they were paying for. Would they have voted the same manner, knowing they would be choosing an AI over an individual? For currently, it is uncertain.

The team states that its research should not be understood as an extreme technocratic proposal to overthrow how wealth is currently redistributed in society. However, it is a research tool that might aid human beings in engineering potentially better solutions than what we have now.

“Our results do not indicate assistance for a form of ‘AI government,’ whereby autonomous agents make policy decisions without human intervention,” the authors write.

“We see Democratic AI as a study methodology for designing potentially beneficial mechanisms, not a recipe for deploying AI in the public sphere.”

The findings are reported in Nature Human Behaviour.

Read the original article on Science Alert.