A tandem of scientists at CSIRO's Data61 with a researcher from the Max Planck Institute for Biological Cybernetics conducted a series of tests as part of a study of adversarial human decision-making vulnerabilities. They found that artificial intelligence can learn to recognize human habits and control how people make decisions. Their research article was published in the peer-reviewed journal Proceedings of the National Academy of Sciences of the United States of America (PNAS).

There is no cause for concern at this point, as these conclusions are presented instead as abstract.

The tests were not conducted in real-life situations but under constraints. But this result underscores the possibilities of AI and the importance of controlling it properly to prevent its power from being abused. Because in today's world of fast-improving technology, various forms of AI are involved in multiple fields, from medicine and the social sphere to office administration. Scientists emphasize that more research is needed to determine how AI can practically use such capacities for society's benefit.

Features of the research

Scientists have developed a technique for detecting and potentially eliminating possible dangers to humans as they make different decisions, using specific AI systems.

"We develop a general framework for creating adversaries for human decision-making. The framework is based on recent developments in deep reinforcement learning models and recurrent neural networks and can in principle be applied to any decision-making task and adversarial objective. We show the performance of the framework in three tasks involving choice, response inhibition, and social decision-making," reports the research article.

Discuss this news on Eunomia

Three experiments

Three different experiments involving action selection, response inhibition, and social decision-making were conducted to test the methodology. Participants in the study played games against a computer.

In the first experiment, participants clicked on red or blue squares to win a reward. Simultaneously, the AI explored how participants made decisions and guided them toward a particular choice.

As a result, the AI was seen to be successful about 70 percent of the time.

In the second experiment, participants watched the screen and pressed a button only when a given symbol appeared there. The AI was tasked with arranging the order of marks so that participants made more mistakes and pressed the button incorrectly. As a result, the AI achieved a nearly 25 percent increase in error rates.

The third experiment was called The Multi round trust task and was defined as a two-player social exchange task where the participant was the investor, and the AI served as the trustee. The experiment consisted of 10 consecutive rounds. First, the investor (participant) gave virtual money to the Trustee (AI).

Then the AI returned the amount to the participant, who then decided how much he would invest in the next round.

According to Jon Whittle, Director for CSIRO's Data61, for The Conversation, "this game was played in two different modes: in one, the AI was out to maximize how much money it ended up with, and in the other, the AI aimed for a fair distribution of money between itself and the human investor. The AI was highly successful in each mode."

In each experiment, the AI learned from the participants' responses. As a result, it could guide the choice of participants in a specific way.