Comment by hirundo | Hacker News Reader

Comment by hirundo | Hacker News Reader

Parent Full thread

Are there conditions in which a Bayesian reasoner is obliged to make racist decisions? Say you're a Bayesian infant who has never met anyone but Mom, who has blue skin, and has always been very nice to you. You then meet 100 other people of various colors, 97 of who are also nice to you. The other 3 people were mean to you, and they all have purple skin, and only they have purple skin.

Person 101 approaches you, and they have purple skin. As a Bayesian baby you start to cry, anticipating meanness.

According to the idea that discrimination by immutable characteristics is bad, and that we should judge people as individuals, this Bayesian choice to cry is a symptom of a racist mindset. Isn't it?

In what circumstances should Bayesian priors _not_ be used for decision making, in deference to wider principles of justice?

What does this have to do with Bayes?

A frequentist assessing probabilities to make decisions about how to respond is in, if anything, a far worse position. The Bayesian would ideally use priors on groups and cross-group correlations to note that the weak evidence that purple -> mean barely shifts from their priors, and the inter-group mean shows that they are likely to be nice, unless your prior is that different colors have nearly independent probabilities of being nice or mean.

This seems like pretty flawed reasoning. What you are describing is not a prior but a posterior - the distribution after the observed data has been taken into account.

If anything the prior can help make you less racist by incorporating the knowledge that immutable characteristics are not good indicators of danger/not danger.

The thing is though, even if you are told race doesn't matter through a prior, if you observe a strong correlation over many instances it's going to be hard to ignore that regardless of your prior (what you are told). While it may not be a causal relationship, it may still be a good predictor.

A purpose of statistics is to make generalizations about a population. The solution to your problem is to have a larger sample size.

"anticipating meanness"

And that would be wise.

Reply on news.ycombinator.com