Research published in the journal iSciencerevealed an unusual effect. It turns out that people tend to behave less fairly towards AI with female, non-binary or neutral characteristics.than to AI designated as male. The work shows that gender stereotypes manifest themselves even in relationships with intangible digital agents.

How the experiment went

Participants were asked to play the classic prisoner’s dilemma game. There are two options available in this game. First –cooperate and gain mutual benefit. Second –try to win at the expense of your partner, taking advantage of his gullibility. If both choose mutual assistance, the result is optimal. If one betrays and the other does not, the traitor gets points. When both play alone, the winnings are minimal.

To test the influence of gender associations, the researchers designated partners with four options. These were “female”, “male”, “non-binary” and “genderless”. They could be both humans and AI agents. This separation helped to see how the behavior of the participants changed.

Who plays fair and who cheats?

The authors of the paper found that people were about 10% likely to use a dishonest strategy when playing with an AI rather than a human. At the same time, participants were willing to cooperate with agents endowed with feminine, non-binary or neutral characteristics than with “male” partners. They assumed that such agents themselves would choose an honest strategy.

Women were likely than others to show trust in “female” agents and were less likely to take risks when playing with male characters. This is reminiscent of the well-known phenomenon of homophily, when a person is willing to interact with those whom he perceives as “similar to himself.”

Why there is a desire to exploit AI

The refusal to cooperate could be explained by two motives.The first is the reluctance to be deceived. The second is an attempt to take advantage of the partner’s perceived honesty.Researchers considered the second option to be a manifestation of exploitation.

It turned out that participants often choose a dishonest strategy towards female and “genderless” agents than towards male ones. If the partner was an AI, the likelihood of exploitation became even higher. Men were likely than women to take advantage of a partner’s weakness, and they also differentiated between humans and AI strongly.

Participants were likely to “exploit” female AI partners because they subconsciously perceived them as cooperative or less able to resist. This reflects existing biases against women.AI exploitation is likely to occur with agents considered vulnerable by participants.

Human expectations shape behavior in AI interactions. People are likely to trust female or neutral AI, but will take the opportunity to benefit if the game suggests such a scenario.

What does this say about the future of AI?

The authors noted that modern AI is becoming and “humanized.” Systems have names, voices, appearances, and gender tags. All this helps people trust machines faster. However, this approach can unintentionally reinforce existing social imbalances.

If people become accustomed to exploiting some agents and avoiding others, this can affect their behavior in the real world. This is especially important to consider in the future, when AI will manage transport, schedule work tasks and become a constant partner in decision making.

“Such distortions can perpetuate bad habits and increase gender discrimination,” the study authors emphasize.

To avoid this, it is important for developers to take such effects into account when creating interfaces and interactions, and strive for fair and neutral AI.