AI learns women discrimination and racial discrimination from human language

In the development of artificial intelligence (AI), it is studied whether input similar to human can be realized by inputting a lot of information. However, with computers that are loaded with a large amount of human-made documents, the problem of precisely reproducing prejudices and discriminatory emotions that exist strictly behind the human language is occurring.

AI Learns Gender and Racial Biases from Language - IEEE Spectrum
http://spectrum.ieee.org/tech-talk/robotics/artificial-intelligence/ai-learns-gender-and-racial-biases-from-language

Machine learning, AI tied not to associate words such as "insects" and "weapons" with "fun things", while AI related to words such as "flower" and "music" is related to "fun things" I know that. Also, there are many cases that "European American names" are more positive than "African American names", and "women" tend to relate "artistic" to "artistic" than "mathematics" It has been confirmed. These are thought to be examples in which AI's thinking reflects stereotypes and discriminatory feelings of human beings who are infesting society.


However, Dr. Arvind Narayanan, a computer scientist at Princeton University says, "In all cases where machine learning contributes to perceptual tasks, machine learning reproduces human prejudice and reproduces what is behind us "There is a concern that it is a concern that it is a concern that it will be reflected in human learning by machine learning, and warning that computers will also conduct prejudice and discriminatory thinking just as human beings are problematic.

Dr. Narayanan and the research group at Bath University have used statistical tests called "Implicit Association Test" which psychologists use when investigating human prejudices to clarify the prejudice that may occur in the natural language learning process, I applied it to the AI ​​system and examined the degree of prejudice of AI. The research group gathered 2.2 million words from the Internet and learned by AI in order to investigate whether prejudice on race and gender was reproduced in AI's thinking. Then, when we developed and investigated a test called Word-Embedding Factual Association Test (WEFAT) which measures the relation with words, the statistical strength of the relevance of machine learning words is " The ratio of women in 50 occupations "has been found to be strongly related.


Dr. Narayanan says about the test result "It is found that the relevance of occupations and sexual words given to women by purely language use alone is 90%", and word relevance and labor statistics It seems that the strong correlation with the document was surprising for the researchers. For this reason, whether the prejudice can be removed from the input information when doing machine learning, etc. is being studied for the necessity of rules in terms of ethics.

The research results of Dr. Narayanan et al. Highlight the realistic problem that AI learns prejudice and discriminatory thinking in the background of words by learning from documents written by humans, By loading it separately and measuring the degree of potential prejudice, it is possible to clarify how social prejudice has been fostered with the passage of time, obtain clues to find out the cause and resolve it It has also been pointed out that it is possible to take advantage of the prejudice that artificial intelligence learned as a powerful tool to do.