プリンストン� �学の研究チー� ��.pdf

Researchers from Princeton University and the University of Bath found that artificial intelligence and machine learning models can perpetuate human biases and prejudices present in language. They used a statistical machine-learning model, GloVe word embedding, trained on a corpus of text from the Web, to replicate standard human biases as exposed by the Implicit Association Test. The results indicate that language itself contains recoverable and accurate imprints of historic biases, which are captured by machine learning along with the rest of semantics. The study has implications for AI, machine learning, psychology, sociology, and human ethics, suggesting that mere exposure to everyday language can account for biases and that prejudice must be addressed as a component of any intelligent system learning from human culture.