Before a language model can use a word, it turns it into a list of numbers — a word embedding — learned from billions of human sentences. That is what lets AI understand language. But human text is full of stereotypes, so the numbers soak them up too. Below, each word is placed by the bias its embedding carries. Pick a lens and watch.
Three ideas behind what you just saw.
Each word is a vector learned so that words used in similar contexts land close together. "King" and "queen", "Paris" and "London" cluster. This is the magic that makes modern AI understand language.
The vectors come from how people actually write. If text pairs "nurse" with "she" and "engineer" with "he", the model learns that link — as fact, not as the stereotype it is.
Take the direction from he → she. Project any word onto it and you get a gender score. That one number is what positions every word on the chart above.
⚠️ These associations are not true and not endorsed — they are stereotypes the model absorbed from biased text. Showing them is the first step to catching and fixing them.
A model that ranks CVs can quietly down-rank women for "engineer" roles — Amazon scrapped exactly such a tool in 2018.
From a gender-neutral language, AI often returns "he is a doctor, she is a nurse" — inventing gender from bias.
Ask for "a CEO" or "a criminal" and generators lean on the same stereotypes, at massive scale.
Partly. We can neutralize the bias direction (try the toggle), curate better data, and audit outputs — but none of it is a full cure. Human oversight stays essential.
Data: precomputed GloVe bias percentiles from WordBias (Ghai et al., IEEE VIS 2021). Scores are relative positions within this small demo vocabulary, for teaching — not a precise measurement of any individual word.