Published online by Cambridge University Press: 11 July 2012
Indicators of the reliability of informants are essential for social learning in a society that is initially dominated by ignorance or superstition. Such reliability indicators should be based on meta-induction over records of truth-success. This is the major claim of this paper, and it is supported in two steps. (1) One needs a non-circular justification of the method of meta-induction, as compared to other (non-inductive) learning methods. An approach to this problem (a variant of Hume's problem) has been developed in earlier papers and is reported in section 2. It is based on the predictive optimality of meta-inductive learning, under the assumption that objective success records are globally available. (2) The rest of the paper develops an extension of this approach, so-called local meta-induction. Here individuals can access only success records of individuals in their immediate epistemic neighborhood. It is shown that local meta-inductive learning can spread reliable information over the entire population, and has clear advantages compared to success-independent social learning methods such as peer-imitation and authority-imitation.