A new study has shown that manifesting prejudice towards others does not require special intelligence and can easily develop in artificially intelligent machines.
Psychologists and information technology specialists at the University of Cardiff and MIT have shown that groups of autonomous machines can exhibit biases by simply defining such behavior, copying it, and mutually teaching it.
It may seem that prejudice is a purely human phenomenon, requiring human intelligence to form opinions or stereotypes about a person or group. While some types of computer algorithms have already exhibited biases such as racism and sexism based on the study of public records and other human-generated data, new work demonstrates AI's ability to self-develop biased groups.
The research is published in Scientific Reports. It is based on computer simulations of how biased virtual agents can form groups and interact with each other. During the simulation, each individual decides whether to help someone from his group or from another, depending on the reputation of that individual, as well as his own strategy, which includes their levels of prejudice towards outsiders. After conducting thousands of simulations, each individual learns new strategies by copying others - whether they are members of their own group or the entire "population".
The relative cumulative frequency of agents' characteristics by the level of prejudice / Roger M. Whitaker.
“After running these simulations thousands and thousands of times in a row, we began to understand how bias develops and what conditions are needed to cultivate or prevent it,” said study co-author Professor Roger Whitaker of the Institute for Crime and Security Research and the School of Computer Science and Computer Science at Cardiff University. “Our simulations show that bias is a powerful force of nature, and through evolution it can be stimulated in virtual populations to harm broader connections with others. Protection against prejudiced groups can inadvertently lead to the formation of other prejudiced groups, provoking greater division of the population. Such widespread prejudice is difficult to reverse.”
The research data also includes individuals who increase their levels of bias by preferentially copying those who get the best short-term results, which in turn means that such decisions do not necessarily require special abilities.
“It is entirely plausible that autonomous machines, capable of identifying with discrimination and copying others, may in the future be susceptible to the phenomena of prejudice that we see in society,” continues Professor Whitaker. “Many AI developments that we see today involve autonomy and self-control, that is, the behavior of devices is also influenced by those around them. Recent examples include transportation and the Internet of Things. Our research provides a theoretical insight into where simulated agents periodically turn to others for resources."
Promotional video:
The researchers also found that under certain conditions, including the presence of more divided subpopulations of the same society, bias is more difficult to reinforce.
“With a large number of subpopulations, unbiased group unions can cooperate without being exploited. It also diminishes their minority status while reducing their susceptibility to establishing bias. However, this also requires circumstances in which agents are more favorably disposed towards interactions outside their group,”concluded Professor Whitaker.
Vladimir Guillen