Bias is a bad word

The meaning of a word is its use in the language.

• Ludwig Wittgenstein, Philosophical Investigations

In the abortion debate, one side calls itself pro-life; the other pro-choice. Each picked the label that sounds principled and left the ugly framing for the opponent: anti-abortion, anti-choice.¹ Pro-life sounds like basic decency. Anti-choice sounds authoritarian.

The words that do the most damage sound reasonable while smuggling in a verdict.

Bias is one of those words.

A Word Doing Too Much Work

In a dictionary “bias” is given three jobs:

1. In statistics, it is a systematic error.
2. In psychology, it is a cognitive shortcut: confirmation bias, anchoring bias, loss-aversion.
3. In everyday speech, it’s prejudice.

A biased coin and a biased jury might share a word, but in reality they share very little.

Yet when we say a machine learning model is “biased,” we borrow from all three meanings at once, and the listener gets to pick the most alarming.

What Models Learn

A language model picks up the statistical regularities of its training text. Tables typically have legs. Birds typically fly. Nurses appear as “she” more often than “he.” (at least currently).

The model has no id to motivate malicious action. It reflects the statistical reality of the text we wrote down.

Caliskan et al. showed that word embeddings trained on ordinary English text reproduce the same associations found in human implicit association tests.² The model “knows” that flowers are more pleasant than insects, and that European-American names are more associated with pleasant words than African-American names. Both are accurate reflections of the training data.³

Use Priors

There is a better word: priors.

In Bayesian statistics, a prior is a belief held before seeing new evidence; the word carries no moral judgement.

When we say a model has priors instead of biases, the conversation shifts from blame to mechanics: where did this prior come from? Is it accurate? Should we update it? The model has priors because we have priors; it learned them from us.

You can still be outraged.

A hiring algorithm that associates “senior engineer” with “male” is harmful. An LLM that defaults to “The criminal was [racial stereotype]” is also harmful.

But calling these “biases” makes the model sound like a bigot when it is a parrot. Much like a child who learns a swear word — it isn’t wrong to learn it; we just wish it hadn’t.

If the model is biased, the fix is moral correction: de-bias it. If the model has priors, the fixes are engineering: identify harmful priors, and update.

Teaching machines is much like teaching ourselves

Perhaps, much like children, we can teach machines to hold better priors than the ones we have. The first step is just a word.

When you want to say “bias,” say “priors” instead.