ScienceDaily: Search results can have a positive impact on our social inequalities

Date:

Share post:


New research by a team psychologist researchers shows that even though internet searches are gender neutral, they still produce male-dominated results. Additionally, the search results may have an effect on users through promoting gender bias or potentially influencing hiring decisions.

The article appears in the journal. Proceedings of the National Academy of Sciences(PNAS) is one of the latest to reveal how artificial intelligence (AI), can alter our perceptions.

“There is growing concern about algorithms used by modern AI system produce discriminatory outputs,” Madalina Vasceanu, a postdoctoral fellow at New York University’s Department of Psychology, and the paper’s principal author, says. Their use by humans could lead to the expansion, not reduction, of existing disparities.

“These findings call to mind a model of ethical AI which combines human psychology and computational and sociological methods to illuminate the formation operation and mitigation of algorithmic biased,” says David Amodio (a professor at NYU’s Department of Psychology) and the University of Amsterdam.

Experts in technology have raised concerns that modern AI systems can produce discriminatory outputs. This could be because they are trained using data with societal biases.

Meredith Broussard (author of): “Certain 1950s views about gender are actually still embedded into our database systems.” Artificial Unintelligence: How Computers Misunderstand Our WorldThe Markup heard from the Arthur L. Carter Journalism Institute Professor at NYU earlier in the year.

Vlasceanu, Amodio believe that AI may lead to more disparities than reduction.

They conducted research to address this possibility.

They first drew on the Global Gender Gap Index, which includes rankings of gender inequality in more than 150 countries. The GGGI is a measure of gender inequalities in economic participation and opportunity as well as in education attainment and survival. Each country also receives societal-level gender inequalities scores.

To assess gender bias in search results or algorithmic output, the team examined whether words that should be used with equal probability to refer to a man and a woman, such “person”, “student” or “human,” were more commonly assumed to be a male. They then used Google to conduct the research. ImageSearches for “person” in a country (in its dominant language) have been performed across 37 countries. The results revealed that searches for “person” in a nation (in its dominant local language) yielded more male images than those with higher gender inequality. This shows that algorithmic gender bias can be linked with societal gender equality.

Three months later, the researchers did the same study again with 52 countries, including 31 of the original study. These results were consistent in the first study. They confirmed that algorithmic output (i.e. internet searches.

Vlasceanu & Amodio then attempted to determine if people’s perceptions, and ultimately their decisions, can be influenced by search-engine results.

They conducted experiments that involved nearly 400 male and female U.S. participants.

Participants were instructed that they would be viewing Google image searches results for four professions they weren’t familiar with, chandler, draper and peruker. To represent the Google image search results of the keyword “person”, for countries with high levels of global gender inequalities (roughly 90% men to 10% in Turkey and Hungary), as well as those with lower global gender inequalities (roughly 50% men or 50% women in Finland and Iceland) from the 52 nation study above, the gender composition of each profession’s images was selected. Researchers were able to replicate the results of different internet searches.

Before they could view the search results, participants made prototypicality judgments about each profession. For example, “Who is more likely a peruker, man or woman?”These were used to establish baseline perceptions. Participants, male and female, rated these professions more likely to be men than women.

But, these questions were answered differently when they were asked. After Viewing the image search results, you can see the participants in the low quality conditions ReversedTheir male-biased prototypes compared to the baseline assessment. However, the high-inequality group retained their male-biased perceptions and reinforced their perceptions of these prototypes.

They then looked at how internet searches might influence hiring decisions. Participants were asked to rate the likelihood of a man or a woman being hired in each field (“What kind of person is most likely be hired as peruker?”). When presented with two job candidates for the same profession, participants were asked to decide which one they would hire.”).

As with other experiments, exposure to images in low-inequality conditions produced more egalitarian judgements of male and female hiring tendencies within a profession as well as a higher probability of selecting a woman job candidate than exposure to images sets in high-inequality.

“These results suggest bias propagation between society and AI,” Vlasceanu, Amodio write. Amodio adds that “findings prove that societal levels are evident in internetsearch algorithms and that exposure can lead human users think and potentially to act in ways which reinforce the societal inequality.”

The NYU Alliance for Public Interest Technology (VICI 016.185.058) and the Netherlands Organization for Scientific Research funded the study.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

spot_img

Related articles

The Benefits of a Creative Subscription with Envato Elements

If you’re looking to find and use royalty-free images and other creative resources, it can be tough to...

Get an Additional ₹100 Cashback When You Pay with Domino’s Digital Wallet Partners

How does ₹100 cashback sound? Find out more about the additional cashback you can get when you pay...

Get Unlimited Access to DataCamp’s Library of Online Courses

DataCamp subscriptions enable access to over 300 courses, as well as projects, assessments, and additional content. Whether you're...

Why You Should Buy from DaMENSCH: The Best in Quality, Service, and Value

Buying products online can be dangerous; you never know if you’re getting an authentic product, or one that’s...