This 2016 paper by Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai showed that word embeddings, the numeric vectors that machine-learning systems use to represent the meaning of words, absorb the social biases present in their training text. The authors worked with embeddings trained on Google News articles, a corpus that powered many real systems at the time.
Their memorable demonstration used the analogy structure that embeddings are famous for. Just as the vectors solve “man is to king as woman is to queen,” the same arithmetic produced “man is to computer programmer as woman is to homemaker,” and linked words like receptionist and nurse more closely to female than male. The bias was not random noise; it was captured by a consistent direction in the vector space, which the authors could isolate and measure.
The paper then proposed a debiasing procedure that neutralizes gender associations for words that should be gender-neutral, such as occupations, while preserving legitimately gendered relationships like queen and king. The method became a reference point, and also a subject of debate, in the larger conversation about whether bias can be engineered out of representations or whether it must be addressed in how systems are built and used.
Why a business reader should care: any model trained on human-generated text inherits the patterns in that text, including stereotypes. If embeddings feed hiring, search, or recommendation systems, those stereotypes can quietly shape who gets surfaced and who gets filtered out.