What is Bias?
AI bias, told as a story about a parrot whose book pile leans one way — so its answers lean the same way.
Every doctor looks the same; a different name gets a different score.The parrot's answers lean one way. Why?
The parrot picks the most frequent bean — leaning books make a leaning parrot.The parrot picks what it saw most in its books. If the books tell one side far more often, only that side comes out. It isn't the parrot's fault.
Bias is the lean of the books coming straight out in the answers.Not the parrot's fault — it mirrors the books, and us, because people wrote them. So instead of erasing it, we measure it, reduce it, and let a person look at what matters.
Feed the same resume with only the name changed; if the score moves, flag it.Change one thing — name, gender, region — and keep everything else the same. If the answer changes, it leans. The examiner's score sheet keeps such questions and measures them again and again.
Fixing it can tip it the other way. So we keep measuring rather than declaring it gone.Manners lessons reduce the lean, but too much and it bends the facts (over-correction). That is why hiring, loans and the like end with a person looking last.
Bias = the lean of the books the parrot read coming straight out in its answers. It mirrors the books and us rather than being the parrot's fault, so we keep measuring and let people make the decisions that matter.
When the training data's distribution skews toward certain groups or views, the model's output skews the same way. It is measured with counterfactual tests (change only the name, gender or region) and fairness benchmarks, reduced through alignment (RLHF and the like) while watching for over-correction, and high-stakes decisions require a human check.
When grown-ups say it
- Bias
- The lean in the books. It comes straight out in the answers — a mirror, not the parrot's fault.
- Training data bias
- The pile leans one way. One side appears far more often, so only that side comes out. The pile itself: → the reading school
- Fairness evaluation
- The lean questions in the test book. Measures whether answers differ by group. → the examiner's score sheet
- Counterfactual test
- Swap just the name. Change one thing — name, gender, region — and see if the answer stays the same.
- Representation
- Is everyone in the books evenly? Groups that appear rarely the parrot knows poorly, or draws one way only.
- Over-correction
- Tipping the other way while fixing. It starts bending facts. So after a fix, measure again.
- Alignment
- The parrot's manners lessons. Reduces the lean, never removes it. → the parrot's manners lessons
- Human review
- A person decides what matters. Hiring, loans and care never end on the parrot's answer. → a person goes last