(ORDO NEWS) — Artificial intelligence (AI) can develop wealth distribution methods more popular than human-designed systems, according to a new study.
Findings made by a group of researchers from a British artificial intelligence company. DeepMind show that machine learning systems are not only good at solving complex problems in physics and biology, but can also help achieve more open social goals, such as the goal of creating a just and prosperous society.
Of course, this is not an easy task. Creating a machine capable of delivering useful outcomes that people actually want called “value matching” in AI research is complicated by the fact that people often disagree about the best way to deal with a wide variety of issues, especially social, economic, and political issues. problems. .
“One of the key barriers to value negotiation is that human society allows for multiple views, making it unclear whose preferences AI should be aligned with,” the researchers explain in a new paper by first author and DeepMind researcher. scientist Raphael Coster.
“For example, political scientists and economists often disagree about what mechanisms will make our society function most fairly or efficiently.”
To bridge the gap, the Researchers developed a wealth distribution agent that had human interactions (both real and virtual) embedded in its training data essentially guiding AI toward human-preferred (and hypothetically fairer overall) outcomes.
While AIs can produce three surprising results, they can also come up with far from desirable social outcomes if left to their own devices; Human feedback can help guide neural networks in the right direction.
“There is a growing awareness in AI research that in order to create human-compatible systems, we need new research methods in which humans and agents interact, and increased efforts to learn values directly from humans to create value-driven AI,” they write. researchers.
In experiments that involved a total of thousands of people, the team’s AI agent, called “Democratic AI”, studied an investment exercise called the public goods game, in which players receive various amounts of money and can contribute their money to a public fund, and then receive income from the fund corresponding to the level of their investment.
In a number of different playstyles, wealth has been redistributed among players through three traditional redistribution paradigms strictly egalitarian, libertarian, and liberal-egalitarian—each of which rewards players’ investment differently.
A fourth method has also been tested. , called Hu Human-Centered Redistribution Engine (HCRM), developed using deep reinforcement learning, using feedback data from both human players and virtual agents designed to mimic human behavior.
Subsequent experiments showed that the HCRM system for paying out money in the game was more popular with players than any of the traditional redistribution standards, and also more popular than the new redistribution systems developed by human judges who were interested in creating popular systems by receiving small voice payments.
“The AI discovered a mechanism that reversed the initial imbalance in wealth, sanctioned free riders, and successfully won a majority of votes,” the researchers explain.
“We show that the same democratic tools for reaching consensus can be used to equalize values that are used in the wider human society to elect representatives, determine public policy, or make judicial decisions.”
It’s worth noting that the researchers admit their system raises a number of questions – basically that the value alignment in their AI revolves around democratic settings, meaning that an agent can actually exacerbate inequality or prejudice in society (provided they’re popular enough, to get the majority of the people to vote for them).
There is also the issue of trust. In the experiments, the players did not know who was behind the wealth redistribution model they were paying for. Would they have voted the same way, knowing they would prefer AI to humans? At the moment this is unclear.
Finally, the team says that their study should not be seen as a radical technocratic proposal aimed at debunking the actual redistribution of wealth in society, but rather a research tool that can help people develop potentially better solutions than those currently available.
“Our results do not imply support for a form of ‘AI government’ in which autonomous agents make political decisions without human intervention,” the authors write.
“We view democratic AI as a research methodology for developing potentially useful mechanisms, not as a recipe for deploying AI in the public sphere.”
Contact us: [email protected]