Numerous studies from the judgment and decision-making literature have
demonstrated that people are prone to systematic errors on probabilistic
reasoning tasks.  On their face, these findings suggest that probabilistic
inference is not a natural human ability.  Rather, people employ shortcut
strategies, such as Tversky and Kahneman's ``representativeness
heuristic'', that follow a logic distinct from probability theory.  This
conclusion contrasts sharply with recent findings in other areas of human
learning and inference, where a number of researchers have constructed
successful Bayesian models of tasks ranging from concept learning,
syllogistic reasoning, and hypothesis testing to causal inference, word
learning and predicting the future.  Are we therefore forced to conclude
that people have a natural ability for probabilistic inference in every
situation *except* when they are explicitly asked to manipulate
probabilities?  In this talk, I will first review some of my work on
Bayesian models of concept learning and then try to reconcile this work
with the conclusions of Tversky and Kahneman.  I will argue that the
representativeness heuristic, rather than being seen as an alternative to
probability theory, is itself best described as a Bayesian computation,
and one that is intimately related to the Bayesian computations underlying
our abilities to learn concepts, words, and causal relations from
examples.  Time permitting, I will also discuss a parallel with Popper's
arguments against probability theory as a viable logic of confirmation in
scientific inquiry.