What causes discrimination? Conventionally, there are two views on this. One is taste discrimination; some people would rather not hire blacks or women or gays or whatever simply because they don't like them. The other is statistical discrimination; employers believe (rightly or not) that a particular group is, on average, less able and so are reluctant to hire from that group.

However, a recent experiment suggests that there's another form of discrimination that is harder to eliminate - a simple lack of Bayesian reasoning.

Researchers first got people to perform some mental rotation tasks, and then split the performers into two groups. In one group, which they called K, 43% were top performers (who got 13 or more of 24 right) and 57% poor performers. In the other group, called L, 14% were top performers and 86% poor ones.

A separate group of people were then given a choice. They could take a small sum of money, or they could gamble on an individual being drawn from group K or L where they would win '20 if the performer were top but nothing if they were not.

Obviously, in this gamble one has a 43% chance of winning if you can choose from group K.

But here's the thing. Subjects were told that the draw from group L was rigged, in such a way that someone would be drawn from group K and if s/he were a top performer, a top performer from L would also be picked. This means that subjects had a 43% chance of winning, whether they drew from group K or L.

However, despite knowing this, subjects preferred to bet on group K. In other words, they paid attention to base rate probabilities (43-57 vs. 14-86) even though they were irrelevant. This is Bayesian conservatism.

This is not taste-based discrimination, because K and L are neutral terms. Nor is it statistical discrimination because subjects' actual odds of success were identical in both groups. It's plain irrational.

This might have nasty implications in labour markets*. It suggests that group labels stick to individuals, even if the particular individual does not have that particular attribute. So, for example, if an employer is loath to hire women for fear they will leave to have children, he might be reluctant to hire even women who could prove they have no such intent; though lesbians earn more than straight women, they earn less than men. Fannie Hurst's line that a woman has to be twice as good as a man to go half as far might, therefore, be true.

In this sense, stereotypes are dangerous not merely because they can be self-fulfilling, but also because they damage the life-chances even of people who do not conform to them.