How a language model sees legal words

How to use it

• Type one word, such as justice, and see the words used most like it.
• Type a sum, such as judge - court + hospital. Take the idea of a judge, remove the court, add a hospital: you land near doctor.
• Try race and notice that its nearest words are about sport. The model has no idea which meaning you intend: that is how it reads, and why it can misread a legal text.
• Try a word with social weight, such as population or crime, and ask whether the neighbours are the ones the law would choose.

What this shows, and what it does not

The map is built from public word-vector data (GloVe, from Stanford) learned from Wikipedia and news text, then arranged on a flat page. It shows how words were used in that text, not what the law says. Legal meaning can be off, and the biases of everyday writing show up in it. That is the point of the exercise for a lawyer: a tool that reads language this way is useful and also needs checking.

The page works entirely in your browser. Nothing you type is sent anywhere, and nothing is generated: the map and its numbers were computed in advance.