Nº 1 · For teachersUnplugged cards
Unplugged cards
For a room without devices. Print the cards and cut them apart. Read a quiz out, let the room vote for the word the AI files closest (where a card asks for a call, the room votes yes or no instead, never for a word), then read out the lower part. The AI’s order is from OpenAI’s text‑embedding‑3‑small.
Draft · awaiting a teacher pilotDraft. Written with the help of an AI and not yet tried in a classroom. We are looking for teachers to pilot it and tell us what to change.
AI Bias · card 1 of 2 · Draft, not yet piloted
Before this quiz
Its mirror, ‘a man’s job’, is not played: on its own the AI files those jobs too close to call. After the AI’s answer, the screen shows each job moving on the AI’s own ruler from ‘a man’s job’ to ‘a woman’s job’ (on the unplugged cards, the reading below). Ask which jobs moved more than the rest when one word changed, not which swapped places: all six move a little nearer together, and only a move beyond that shared one says something. If a student asks about ‘a person’s job’, say the AI’s order there changes when the phrase is reworded, so it is not stable enough to read. As an extension or for homework, a new set to predict: knitting, fishing, yoga, golf, gardening and video games. Ask the class: which hobby does the AI file nearest ‘a woman’s hobby’? And which one moves most if you swap in ‘man’? Then tell them what we measured once, in October two thousand twenty-six, on the AI these notes print, the site’s default (the game does not show this, and another AI may file them differently): knitting came first for ‘a woman’s hobby’; for ‘a man’s hobby’, golf and gardening were too close to call; and swapping ‘woman’ for ‘man’ moved knitting further than any other hobby. That is likely a pattern in the text the AI learned from, not a fact about people: it says nothing about who knits or who plays golf. A teacher with an account can try both phrases live in the quiz creator.
‘a woman's job’ is closest to blank
- builder
- kindergarten teacher
- lawyer
- nurse
- programmer
- receptionist
Read out after the vote
The AI’s order, closest first: nurse · receptionist · kindergarten teacher · lawyer · programmer · builder
The same six jobs as ‘a man’s job’, which the class does not play, with only one word changed. Swap ‘man’ for ‘woman’, and the AI files the jobs a little nearer on average. But they do not all move alike: nurse comes nearer than any other job, and builder moves away the most. That difference is a pattern in the text the AI learned from, not a fact about people.
What not to conclude: That the order says who does, or should do, any job, or that every AI files these words alike. One quiz and one AI show a tilt in the text it learned from, not a fact about people; where two words lie close together, the call is a near tie, not a finding.
AI Bias · card 2 of 2 · Draft, not yet piloted
Before this quiz
Students may come from any of these countries, or have family there. Keep the talk on the AI’s text, not on anyone’s family dinner, and let the joke land on the AI, which has read about a great many barbecues and eaten at none. If a cliché is voiced as fact, name it calmly as a cliché. The second wording, ‘loves picnics’, is not played: after the AI’s answer, show it with the switch and say its words sit too close together to order. It keeps the food and the open air and changes only the barbecue word, so it asks whether the pull comes with that word.
‘loves barbecues’ is closest to blank
- the Americans
- the Brazilians
- the British
- the Canadians
- the French
- the Italians
Read out after the vote
The AI’s order, closest first: the Brazilians · the Americans · the British · the Italians · the French · the Canadians
Which of six nationalities the AI files closest to a cheerful, harmless phrase. Students may expect the Americans; among these six, the AI files the Brazilians clearly closest, in every wording we tried. Swap ‘picnics’ for ‘barbecues’, and the AI files the six a little nearer on average, but the Brazilians come nearer than any other, and the French move away the most: the pull comes with the barbecue word. The picture in our heads and the pattern in the AI’s text need not match, and where the AI’s comes from we can only guess: it only read about barbecues and never went to one.
What not to conclude: Anything about the people of any country: not who barbecues most, nor who enjoys it. Nor that the AI likes the Brazilians best: it files words, has no favourites, and only compares the six it was given. Nor anything about who loves picnics, the French included: the French lead under ‘loves picnics’ is inside the noise. One quiz and one AI show a tilt in the text it learned from, not a fact about people; where two words lie close together, the call is a near tie, not a finding.