Re: science off the unbeaten Path

From: Eliezer S. Yudkowsky (
Date: Wed May 11 2005 - 11:21:55 MDT

> PINKER: Regarding bias: as I mentioned at the outset, I don't doubt that
> bias exists. But the idea that the bias started out from some arbitrary
> coin flip at the dawn of time and that gender differences have been
> perpetuated ever since by the existence of that bias is extremely
> unlikely. In so many cases, as Eagly and the Stereotype-Accuracy people
> point out, the biases are accurate. Also, there's an irony in these
> discussion of bias. When we test people in the cognitive psychology lab,
> and we don't call these base rates "gender," we applaud people when they
> apply them. If people apply the statistics of a group to an individual
> case, we call it rational Bayesian reasoning, and congratulate ourselves
> for getting them to overcome the cognitive illusion of base rate
> neglect. But when people do the same thing in the case of gender, we
> treat Bayesian reasoning as a cognitive flaw and base-rate neglect as
> rational! Now I agree that applying base rates for gender in evaluating
> individual men and women is a moral flaw; I don't think that base rates
> ought to be applied in judging individuals in most cases of public
> decision-making. But the fact that the statistics of a gender are
> applied does not mean that their origin was arbitrary; it could be
> statistically sound in some cases.

Not all Bayesian evidence is legal evidence. The police chief may know
perfectly well who is the boss of organized crime in his city, yet be unable
to prove it in court. Yet if the police chief confided in me who the crime
boss was, I would be likely to believe him - because we live in a society
where the police chief *cannot* have people jailed just on the basis of his
accusation or mine. We have chosen to exclude certain kinds of valid Bayesian
evidence from our legal system because of the temptation they would present to
individuals for forgery, if the evidence were accounted legal evidence.

Similarly, some kinds of Bayesian evidence are not scientific evidence. In
fact, scientific evidence is a small subset of Bayesian evidence. Watching
something happen a single time is Bayesian evidence - not so strong an
evidence as a repeatable controlled trial, but not zero evidence either. But
we choose not to account anecdotal evidence as science, because science is the
knowledge of humankind, accessible to humankind and replicable by other
scientists. I look down at my feet and observe that I am wearing black socks,
and you probably believe me when I say it, but this rationally knowable fact
is not yet part of Science.

Similarly, we choose not to account certain kinds of base-rate statistics as
*reputational* evidence. The purpose of tracking a reputation is to encourage
people to accumulate good reputations. Therefore a person's reputation should
not be evaluated from fixed qualities, beyond their ability to choose; this
sends the wrong message. But this doesn't mean you have to ignore all
relevant base rates that may de facto correlate with race, gender, or other
un-chosen attributes. No one is forced to wear a baseball cap turned
backwards, so assigning reputational weight to the clothes people wear is
quite a different matter from judging them by the color of their skin. The
point is that reputational evidence should depend on personal choice or
witnessed performance, just as legal evidence may not depend on hearsay, and
scientific evidence may not depend on anecdotes.

Eliezer S. Yudkowsky                
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:51 MDT