Anyone who tries to use gender-neutral language often uses gender-neutral terms such as “persons”, “people” or “people”. But how gender neutral are these terms really? An analysis of more than 630 billion English-language words from the Internet has shown that even the seemingly neutral terms are mostly used in a context associated with male. This suggests that when we hear terms like “people” or “person”, we tend to think of men rather than women.
What are the implicit concepts behind the words we use? In order to find out, analyzes with the help of artificial intelligence have become established. The artificial intelligence evaluates which words are used in a similar context and therefore probably have a similar meaning. For example, the sentence “Joe puts on the balak to boil water for tea” would imply that “balak” has the same meaning as “kettle”. Likewise, such analyzes have already shown that, for example, there is a high correspondence between the words “scientist” and “researcher” and a higher correspondence of these words with the word “clever” than with the word “instead”.
How gender-equitable are neutral terms?
“Many forms of bias have been studied in this way, such as the tendency to associate ‘science’ with men rather than women,” explains April Bailey of New York University. “On the other hand, there has hardly been any work on how we see a ‘person’.” She and her team have now dealt with this question. To do this, they used a data set of over 630 billion English-language words used on almost three billion websites.
In three sub-studies, Bailey and her team tested the correspondence of neutral terms such as “people” with gender-specific terms such as “men” and “women” and their Agreement with masculine or feminine associated adjectives and verbs. “Our results show that even when we use gender-neutral terms, we prefer men over women,” said co-author Adina Williams of Facebook Artificial Intelligence Research in New York.
“People” with masculine characteristics
The first sub-study showed that words such as “people”, “person”, “human” and “somebody” are used in a similar context to clearly male gender assignments such as “man”, “men”, “male” and “Hey”. On the other hand, there was significantly less agreement on female gender assignments such as “Woman”, “Women”, “Female” and “She”. “This was also true when we excluded the word ‘man’, which can also be used in general for ‘human’, from the analysis,” reports the author team.
For the second sub-study, they focused on adjectives associated with specific gender stereotypes, such as “empathetic,” “family-oriented,” and “friendly” for women and “energetic,” “rational,” and “controlling” for men. Using machine similarity analysis, they proved that these attributions were in fact frequently used in the corresponding gender contexts. Now they analyzed the similarity between the male and female associated adjectives with gender-neutral terms such as “People” – and indeed: Here, too, it was shown that “People” is more likely to be used with male-associated adjectives than with female-associated ones.
men preferred
The third sub-study, which used verbs instead of adjectives (including “admire”, “complain”, “kiss” for women and “honour”, “respect”, “kill” for men) came to the same conclusion. “These results show that the authors of the analyzed Internet texts write (and to some extent presumably also think) more similarly about people and men than about people and women,” the research team concludes. “This suggests that the collective notion of humans favors males over females.”
From the point of view of Bailey’s colleague Andrei Cimpian, this is questionable from a social point of view. “The ideas of ‘people’ form the basis of many social decisions and political measures,” he says. “Since males and females each make up about half of our species, our collective notion of a ‘person’ favoring males leads to unequal treatment of women in decisions based on that notion.” As culture and collective thinking change influence each other, it is important to be aware of this unequal treatment and to take measures, for example when programming future artificial intelligence, to avoid such a distortion.
Source: April Bailey (New York University, USA) et al., Science Advances, doi: 10.1126/sciadv.abm2463