It has been argued that human language is the only learned symbolic communication system in the natural world. Recent computational models (known as "iterated learning models") investigate what sorts of learning mechanism must be in place in order to ensure that such communication systems are culturally stable. There appear to be two key learning biases that are required. Existing models may be reinterpreted in terms of these two biases, their common elements emphasised and related to known properties of child language learning.

The two biases can be expressed in terms of a production function, p(m), mapping meanings to signals, and a reception function, r(s), mapping signals to meanings. Essentially, learning mechanisms that result in culturally stable systems are biased in favour of an injective p(m) which is a subset of r(s).

Problematically, these biases also make incorrect predictions of the structure of human languages. While they correctly predict a lack of synonymy, they also predict few homonyms, which are actually rife in linguistic systems. The reasons for this and modifications to the iterated learning model involving the introduction of syntactic contexts will be discussed.