Godfrey-Smith. The Baldwin effect purports that acquired traits can become genetically assimilated. This is supposedly because learning is fallible and expensive in terms of time and cognitive resources, so a hard-wiring is beneficial. Two objections arise. First, the process seems self-defeating, since it is triggered by learning but ends up giving up plasticity for rigidity. Second, the learning step seems superfluous: why not go straight to the innate trait? To this latter objection three replies have been given. (1) Breathing space (Baldwin). Learning enables the species to survive long enough for the right combination of genes or mutation to occur. This assumes that the ability to acquire a trait is more prevalent than having a gene for it, so the self-defeating argument applies. Also, "do-or-die" situations are very rare. (2) Good learners are genetically closer to the innate configuration for the trait, bringing it within genetic reach. The Hinton and Nowlan simulation which is supposed to show the feasibility of the Baldwin effect actually has this dubious assumption built in to it. The genome is modelled as a string made up of the symbols 0, 1, ?, where 111... is the optimum, and ?-bits leave room for trial-and-error learning. Obviously, good learners (i.e., those who hit upon 111... by randomising their ?s) will generally have more hard-wired 1s than poor learners. So selection for learning and selection for innateness coincide: the model does not leave room for other forms of learning improvement such as skewing the randomisation or increasing the number of trials. (Dennet tries to defend the Hinton and Nowlan model in his chapter, but his argument shows only that selection for learning will bring organisms "closer in learning space" to the desired behavioural trait; this does nothing to resolve the issue of whether this simultaneously means that the organisms will come closer in genetic space to hard-wiring the trait.) (3) Niche construction. The gene is useless before learning. See Deacon below.
Deacon. The niche-construction point of view "breaks the pseudo-Lamarckian mould," for it does not maintain that acquired traits become genetically assimilated. This is a good thing because this simplistic version of the Baldwin effect fares poorly when applied to the evolution of human language (e.g., Pinker). It assumes, implausibly, the ability of incremental genetic changes to reach this predescribed goal; and is also empirically unsuccessful (e.g., innate control of vocalisation has actually decreased in human evolution). Instead one should think in terms of masking. Most genes are masked from selection since they have no substantial phenotypic effect in the current environment. A change in the environment will unmask some of these and mask others. Thus natural selection need not wait for and does not look for the genetic equivalents of learned behaviours; instead learning unmasks a pool of previously silent genes from which support for the learned behaviour can be drawn in a number of ways. For example, when Waddington exposed flies to heat he unmasked a diverse set of previously silent genes scattered in the population, the beneficial ones of which were driven to genetic fixation. The evolution of human language is probably of this type, with the added complexity that the adaptations alters the behaviour in turn.