PNS: "There is natural variation of (heritable) traits, and, from that (heritable) variation, it is the relative success of the creatures that inherit the traits, in surviving and reproducing in a particular environment, that determines which traits are passed on to the next generation (in much the same way that, in Artificial Selection, the animal breeder selects for the traits he prefers)." Notice that there has been no need at all to mention either "Natural Selection," which is just a metaphor, nor "fitness," which really just means "the relative success of the heritable variants in a given environment."
Success in survival/reproduction -- determined by the effects of the environment on the distribution of heritable traits in the next generation -- is what replaces the intentional choices of the selective animal breeder by a mindless process.
More generally, the case of the effects of the intentional choices of selective breeders is just a very special -- and until very recently, highly atypical -- case of this same, general, mindless process, namely, the transmission success of heritable traits being determined by the causal contingencies of the environment in which they occur: Mindful animal breeding is just one of those environments. Moreover, mindfulness itself is just one -- or several -- of those evolved, heritable traits: It is often the traits of one organism that constitute part of the environment of another organism, whether within or between species.
(4) Fodor's criticism of this principle was that artificial selection does indeed work this way: the breeder selects the traits he has in mind, and if they also happen to be correlated with other traits, we can find out exactly which traits he was actually selecting for by asking the breeder which one(s) he was selecting for. (5) But with natural selection we do not have this, because no one has any traits "in mind." So it makes no sense to say that there was something (namely, increased "fitness") that natural selection somehow "selected for," because "selecting for" is something minds do, intentionally, and "natural selection" has no mind. (6) So the theory of natural selection is wrong. This is the wrong conclusion to draw.
The Darwinian mechanism of adaptation -- with blind variation of heritable traits in one generation, and then reproductive transmission-success in the given environment determining the distribution of those traits in the next generation -- is an original, ingenious (and true!) explanation not only of how and why creatures have the (heritable) traits they have, but of the relation between evolution in general, and human animal breeding in particular: The distribution of heritable traits is always determined by how successful they make their bearers in surviving and reproducing, but in the special case of human-bred animals it happens to be conscious humans that are "culling out" the creatures with the "maladaptive" traits (rather than, say, hungry predators doing it).
(7) Not only wrong, but empty, since PNS does not predict or explain what will happen in any given situation: We have to look at the actual history in any given case, and then come up with a (possibly true, but post-hoc) explanation of the outcome in that particular case. It's certainly true that, apart from the original, ingenious (and true!) explanation of how heritable traits evolve in general, Darwin's PNS "merely" provides a methodology for investigating what happened in particular cases (particular traits, particular creatures, particular environments).
(8) And in order to find out which (of potentially very many) correlated traits were the ones that actually caused the organisms that had them to survive and reproduce better, the biologist has to do further experiments and simulations – something that "natural selection" itself did not do, and has no way of doing. The only thing evolution "does" is change the trait distribution from generation to generation as a function of environmental contingencies and their effects on survival/reproductive success. If there are correlated traits, where one of them is the causally effective one and the others just happen to be coupled with it, evolution could not "care" less (until and unless some of the correlated traits turn into a disadvantage for survival and reproduction).
Unlike the calculus of variations or even ordinary engineering, evolution does not optimize; it does not even wield Occam's Razor (at least not in this particular respect): As Herbert Simon (1996) put it, evolution merely "provisionally 'satisfices'" (which just means -- let us hasten to point out, lest someone is again tempted to give it a mentalistic interpretation! -- that the winners, from the prior generation, just have to be better than the current competition, not all possible competition; hence no "counterfactual-supportingness" here either! “Fitness” is local, provisional, relative and approximate, not global, permanent, absolute, or exact, like F=ma).
If we are interested in whether it is the fox's swiftness or its (correlated) smooth tail that is the cause of its adaptive success, we will have to manipulate experimentally to find out (as some experimental as well as computational biologists do).
(9) So evolutionary explanation is really just post-hoc historical explanation; there is no underlying "covering law," and Darwin's notion of "natural selection" is tautological, mentalistic, and explains nothing. Tautological, yes, to a degree. But PNS is not the first tautology to be discovered, and to prove productive, having previously not been known, or noticed. (The law of the excluded middle is perhaps another.) And PNS is not a pure tautology (which would have taken the form: “whatever happens, happens”). There is the restriction to heritable traits (rather than all traits), which led, amongst other things, to the discovery of the mechanism of their heritability (genes and DNA).
But, yes, to say that "the distribution of heritable traits in generation N is determined by the inheritance success of the distribution of heritable traits in generation N-1" is tautological; it's just that this (true) tautology had never occurred to anyone before Darwin. They either thought that traits came from God, or that they had always been there and simply varied randomly.
There is nothing mentalistic whatsoever here; insead we have the mechanical key to the explanation of the origin and diversity of all heritable traits.
(10) Fodor also pointed out that much the same thing is true (and for much the same reasons) not just of evolution but of "learning" -- in particular, Skinner's reinforcement learning. It is true (and has been pointed out by many others too) that both Skinner's reinforcement learning and Darwin's adaptive evolution are based on random-variation plus differential-retention-by-consequences (Catania & Harnad 1988).
But the analogy ends there. Skinner tried to explain the "shape" of all behavior as having been determined by random-variation (in responses emitted by the organism) plus reward ("selection by consequences"). Skinner even has an (implicit and unintended) analogy with animal breeders' artificial selection vs "natural selection": the analogy between animal trainers' "artificial reinforcement" (for doing the performance tricks) and the "natural reinforcement" in our (and other animals') natural lives that "rewards and trains us" to perform as we do.
But most natural behavior is not explained or explainable as the result of shaping by Skinnerian reinforcement. In particular, the (universal) grammatical "shape" of language (UG) is not learned or learnable by children through reinforcement training, because of the "poverty-of-the-stimulus": What the child hears and says, and what corrections (reinforcements) it receives during its brief language-learning period, are demonstrably too underdetermined to be able to learn UG from -- either via Skinner reinforcement learning or via any other learning mechanism. There just are not enough data (especially negative data: errors and error-corrections), or time (Chomsky 1980).
The same, however, is not true of the evolution of heritable biological traits: For the evolution of today's organisms' heritable traits (their gene pool) there has been plenty of time, and of positive and negative data, in the 4 billion years since the "primal soup" (Dawkins 1976). In other words, there is no evolutionary "poverty-of-the-stimulus" problem with the evolution or evolvability of heritable traits, through random variation plus retention as determined by their adaptive consequences. (My guess is that Fodor, in his heart of hearts, thinks that there is such a poverty-of-the-stimulus problem with the evolution of biological traits, and that that is what motivates his scepticism about both learning and evolution! Fodor & Pylyshyn 1988)
Four billion years is plenty more time and data than the child's c. 4 language-learning years for evolving all heritable traits. It is not clear, though, that even that is enough time or data to evolve UG! But that's a very special case. Chomsky (1959) could have refuted Skinner (and did) with a lot more than just that: Reinforcement learning does not explain the "shape" of most nontrivial behavior, probably because a lot of human and animal behavior is not the result of reinforcement learning during the lifetime of the organism. Some of it evolved earlier; and some of human behavior is taught, or self-taught, via reasoning. Besides, even behaviors that are learnable via reinforcement learning are not explained until one provides the internal design of the learner that is capable of learning them (via reinforcement learning): Some behavior may indeed be learnable through trial-and-error-correction alone, but how do we have to be built in order to be capable of learning it through trial-and-error-correction alone?
This calls for the kind of computational and robotic modeling of our total performance capacity that Turing (1950) called for (and that Skinner's impoverished, a-theoretical "reinforcement schedules" certainly did not and could not provide; Harnad 1996). But that does not mean that there aren't mechanisms for general trial-and-error-correction-based learning. (That is what a lot of computational learning theory, including neural net learning, is all about; Harnad 2009). Since no one yet knows the scope of general learning theory, no one is in a position to say what it can or cannot do, except in special cases where you do have a poverty-of-the-stimulus argument -- which no one does have, either for human learning ability in general, or for heritable traits in general, but only for the single special case of UG.
(11) When an organism is rewarded by the trainer for doing one thing (choosing a green triangle) rather than another (choosing something other than a green triangle), it is not even clear -- without further experiments -- what the animal is choosing from among these correlated features (something that's both green and a triangle, or anything that's green, or anything that's triangular, or...), just as it is unclear in evolution which of multiple correlated traits is the adaptive trait. This is certainly true. So in the case of learning, you do the further experiments to see which of the correlated features the animal is actually using; and in the case of evolution, you do the further experiments to see which of the correlated traits are actually causing the reproductive/survival advantage.
This is part of doing the actual science. It in no way impugns Darwin with having had to make illicit use of any nonexistent “mindfulness” in any way. It does not even impugn Skinner with that. It is merely another flaw in Skinner that, besides being oblivious to the limits on what can be explained as being just the outcome of reinforcement learning, he ignored the fact that people have minds and often do things because they've made up their minds to do them. Of course, that fact is not an explanation either. But neither is Skinnerian reinforcement an explanation of the nature or origins of our mental powers. (Darwinian evolution, on the other hand, is very likely to be!)
But, in any case, experimentally unbundling and testing correlated variables is not a problem or handicap to either Skinner or Darwin. It is perfectly normal science.