Kolen and Pollack’s paper makes an excellent point: the level of behavioural complexity that a physical system exhibits is determined in part by how one chooses to observe the system’s behaviour. By focusing on the classification of behavioural complexity in terms of the generative classes of formally specified grammatical models, Kolen and Pollack are able to derive by thought experiments and numerical simulations the nice result that generative class is not a natural, uniquely-valued description of a system. Whether the behaviour of a system appears context-free or context-sensitive, for example, depends on how the behaviour of the system is observed. This result is particularly valuable given the seriousness with which claims about the generative classes of behaviours of physical systems are often taken, not only in descriptive linguistics, but also, for example, in molecular biology (Collado-Vides 1991, Searls 1993). © 1995 Taylor and Francis Ltd.