Programs like Google Translate has experienced a dramatic hike in the ability of language interpretation in past few years. Thanks to the new machine learning techniques and of course to the numerous online text data. It is on these data that the algorithms can be trained.
Everyday machines are acquiring the human-like abilities of language. Along with it, they are also inheriting the deeply ingrained biases hidden in the patterns of language. AI has been seen exhibiting a striking gender and racial biases. Many studies have proved that AI is biased towards gender and races. Is it really biased?
Is AI Biased Or We Are?
People say that this shows AI is prejudiced and biased. No, it shows we are biased and prejudiced and AI is learning it. Because we are assisting AI in its learning. It is reinforced in AI because algorithms are unequipped in consciously counteracting the learned biases, unlike humans.
Machines learn from Word Embedding. This process has transformed the way computers interpret text and speech. This method has successfully helped computers in making sense of the language in past few years.
Word Embedding builds up a mathematical representation of language. The meaning of a word in it is distilled into a series of numbers, called Word Vector and based on which related words and terms words frequently appear together.
For example, in the mathematical language space, words for flowers and pleasantness are clustered closer while words for insects and unpleasantness appear close together. Thus, to AI flower is connected to pleasantness and insect to unpleasantness.
This purely statistical approach has captured the social and cultural context of words differently than a dictionary. Some of these implicit biases in the experiments of human psychology has been captured by these algorithms. Words like female and women became closely associated with arts, humanities, and home while the words male and the man appeared closer to the professions like Math and engineering.