Machine learning applied to ordinary English can identify content corresponding to human biases, according to a new study. Aylin Caliskan of Princeton University and colleagues used a word-embedding method that represents each word in a 2.2 million-word case-sensitive vocabulary as a vector with 300 semantic dimensions, derived from 840 billion “tokens” obtained from crawling the web. The method was able to distinguish human attitudes including likes, dislikes and stereotypes, potentially adding a new “comprehension” skill to computers. The result could lead to new understanding and detection of (perhaps unintended) discrimination based on factors such as race, gender, age or ethnicity.