Caliskan, Aylin, Joanna J. Bryson, and Arvind Narayanan (2017) show that machine learning models, such as GloVe word embeddings, inherit human-like biases from language corpora. These biases reflect historical prejudices, including those related to race, gender, and other topics. They replicate human biases using the Word Embedding Association Test (WEAT) and the Word Embedding Factual Association Test (WEFAT), demonstrating that language models capture semantic associations that align with human stereotypes. For example, European-American names are more associated with positive terms than African-American names, and gender stereotypes are reflected in word embeddings. The study also shows that word embeddings can predict real-world data, such as the gender distribution in occupations, with high accuracy. The findings highlight that biases are embedded in language and cannot be eliminated solely through algorithmic means. The research has implications for AI, psychology, sociology, and ethics, as it shows that exposure to language can perpetuate biases. The authors argue that addressing bias in AI requires more than transparency and diversity; it also involves understanding how language shapes meaning and prejudice. The study underscores the need for careful consideration of biases in AI systems and the importance of developing methods to mitigate them.Caliskan, Aylin, Joanna J. Bryson, and Arvind Narayanan (2017) show that machine learning models, such as GloVe word embeddings, inherit human-like biases from language corpora. These biases reflect historical prejudices, including those related to race, gender, and other topics. They replicate human biases using the Word Embedding Association Test (WEAT) and the Word Embedding Factual Association Test (WEFAT), demonstrating that language models capture semantic associations that align with human stereotypes. For example, European-American names are more associated with positive terms than African-American names, and gender stereotypes are reflected in word embeddings. The study also shows that word embeddings can predict real-world data, such as the gender distribution in occupations, with high accuracy. The findings highlight that biases are embedded in language and cannot be eliminated solely through algorithmic means. The research has implications for AI, psychology, sociology, and ethics, as it shows that exposure to language can perpetuate biases. The authors argue that addressing bias in AI requires more than transparency and diversity; it also involves understanding how language shapes meaning and prejudice. The study underscores the need for careful consideration of biases in AI systems and the importance of developing methods to mitigate them.