Artificial intelligence would be better than humans in the distribution of wealth

Artificial intelligence (AI) can devise methods of wealth distribution fairer than systems designed by people, new research published in Human behavior of nature.

The results, made by a team of researchers from the artificial intelligence company deep mindbased in the UK, show that machine learning systems are not only effective in solving complex problems in physics and biology, but can also help achieve social goalshow to have a just and prosperous society.

Of course, it is not an easy task. It’s called building a machine that can deliver the beneficial results humans really want “alignment of values” in artificial intelligence research, it’s complicated by the fact that people often disagree on the best way to solve all sorts of things, and especially social, economic and political problems. .

“A key obstacle to the alignment of values ​​is that human society admits a plurality of viewpoints, therefore it is not clear which preferences The AI ​​needs to be aligned, ”explains lead author of the study, Raphael Koster.

“For example, political scientists and economists often disagree what mechanisms will make our societies work fairer or more efficient, “added the scientist.

To help bridge the gap, the researchers developed an agent for the distribution of wealth that he had people’s interactions (both real and virtual) integrated into his training data; in fact, it drove AI to human-preferred (and hypothetically more equitable overall) outcomes.

While AIs can produce truly amazing results, they can also reach social conclusions that they are undesirable if left to fend for themselves; Human feedback can help guide neural networks in a better direction.

“In artificial intelligence research, there is a growing understanding that in order to build systems compatible with humans, we need new research methods in which humans and agents interact and a greater effort to learn values ​​directly from humans to build an AI aligned to values ​​”.

In experiments involving thousands of human participants, the team’s artificial intelligence agent, called “Democratic AI,” studied an investment exercise called the public goods game, where players receive varying amounts of money and can contribute their money to a public fund. and then you take a return from the fund corresponding to your investment level.

In different playing styles, wealth has been redistributed to players through three paradigms traditional redistribution strategies: strict egalitarian, libertarian and egalitarian liberal, each of which rewards player investments in different ways.

Also A fourth method was triedcalled Human-Centered Redistribution Mechanism (HCRM), developed using deep reinforcement learning, with feedback data from human players and virtual agents designed to mimic human behavior.

Subsequent experiments have shown that the HCRM system for in-game money payment was more popular with players of any standard traditional redistribution systems, and even more popular than the new redistribution systems designed by human arbitrators who have been incentivized to create systems.

“The AI discovered a mechanism that corrected the imbalance initial wealth, free users sanctioned and successfully obtained the majority vote, “explain the researchers.

“We show it you can use the same tools for aligning the values democratic consensus-building processes used in the broader human society to elect representatives, decide public policies or make legal judgments.

It is worth noting that researchers recognize that their system raises a number of questions, mainly that the alignment of value in their AI revolves around democratic determinations, which means that the agent actually it could exacerbate inequalities or prejudices in society (as long as they are popular enough to be voted for by most people).

and also the question of trust. In the experiments, players were unaware of the identity behind the wealth redistribution model they were paying for. Would they have voted the same way, knowing they would have preferred artificial intelligence to one person? For now it is not clear.

Finally, the team says their research is not to be interpreted as a radical technocratic proposition to reverse the way wealth is effectively redistributed in society, but is a research tool that could help humans to design potentially better solutions than we have now.

“Our results they do not imply support for some form of “AI governance”in which autonomous agents make political decisions without human intervention, “the authors write.

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

Source: ScienceAlert