In Davidson's response ("The Social Aspect of Language") to Michael Dummett's criticism of Davidson's 1986 article "A Nice Derangement of Epitaphs," he says that "what bothers Michael is [...] my failure to appreciate that the concept of a speaker meaning something by what he says depends on the notion of a shared language and not the other way around" (Truth, Language, and History, p. 111). Naturally I agree with Davidson here; but I do have a few concerns about some of the sub-morals to be drawn. I intend to talk about those concerns eventually, but first let me deal with a broader issue, which is that Dummett doesn't seem to have any idea what Davidson is talking about, something which (as you can imagine) renders his criticism somewhat ineffective.
Apparently Dummett thinks that in speaking this way Davidson violates Wittgensteinian strictures against "private languages". But an idiolect isn't the same thing as a "private language" at all. Wittgenstein's target in those famous sections of PI is the Cartesian idea that one can fix the meaning of one's own words by a form of "inner ostension" – that I can as it were "point" to some "inner" mental item and say "when I say X I mean that." This is a fairly specific manifestation of the more general Cartesian picture which has been Wittgenstein's target from the beginning of the book.
As an aside: this can help explain a strange phenomenon in contemporary attitudes toward Philosophical Investigations. Most analytic philosophers who deal with Wittgenstein at all regard the first quarter or so of the book as nothing more than throat-clearing and hand-waving. That's why Kripke's book had such an impact. It said: when most people think of PI, they think of the Private Language Argument. But there's some stuff before that (i.e., the rule-following considerations), of which the PLA is just a specific instance! Well, yes (duh); but with that in mind, perhaps we might keep going back before §142 (imagine that) to find the real core of the book. On this latter reading, the PLA, while interesting, in one sense doesn't really tell us anything we couldn't already have guessed. The book's real subject is the more general (and deep-seated, so much so as to be virtually invisible) Cartesian attitude, and what it takes to render it both visible and treatable at the same time (which turns out not to be as easy as it sounds, as the two tend to get in each other's way).
Now Davidson doesn't make as big a deal about his anti-Cartesianism as Rorty does (his own or Davidson's), which is ironic as Davidson's is the more effective version. But in any case, it would surely be odd for Davidson to set up his entire interpretive system as he does specifically to avoid the Cartesian "inner" – and then fail to notice that he falls into what by that point in PI is a fairly straightforward manifestation of that idea.
But of course he doesn't do this. "Idiolect" is Davidson's term for that structured set of linguistic dispositions attributed by an interpreter to a particular person at a particular time and place. The basis for these attributions, in Davidson's account, is of course the interpreter's observations of, and interactions with, the interlocutor in question, over a period of time. It is not new to "Nice Derangement," but goes back to "Radical Interpretation" and other mid-70's papers, that such attributions of a person's meanings cannot be delivered independently of attributing beliefs to him at the same time – and this requires shared interactions with an objective world. There is no question of meaning's dependence on a purely subjective "inner."
The worry about "inner" ostension of meanings was the typically Cartesian one that for all we know from the "outside," someone might mean something entirely different from the meaning we attribute to him on the basis of his verbal and physical behavior (and our own understanding of our shared environment). Dummett's criticism amounts to the charge that in making the idea of a "shared language" dependent on attributions of meaning achievable without previous agreement (i.e. "linguistic conventions"), Davidson leaves open a very real possibility of attributing to a speaker some meanings he had not "agreed to" and might therefore have his own ("internal"?) ideas about. Or something – I don't even see room for such criticism here, but it must be something like that or the PLA couldn't come up at all.
For this is exactly wrong. The whole point of "Nice Derangement" is to account for the manifest success of communication and understanding, even in cases, such as malapropisms, where such success cannot be accounted for by the traditional model (of previously established linguistic conventions). Of course, in any particular case, you may simply deny that understanding has indeed occurred – just as you may feel obliged to say of any of my beliefs that they "might be false"; but part of Davidson's point is that such skepticism about meaning would manifest exactly the sort of theoretically-driven perversity as does Cartesian skepticism about belief.
In any case an "idiolect" is precisely not a "private language." In attributing meanings to a speaker, I thereby indicate that they are shared: we have used his language to communicate. In this sense, defining a "sociolect" such as English or Flemish is, as Davidson elaborates Dummett's complaint, "the philosophically rather unimportant task of grouping idiolects". Naturally languages of this sort are "shared"; but at the more fundamental level, the sharing in question is not at all dependent on the sort of "linguistic conventions" one uses to make the broader, relatively (conceptually!) straightforwardly empirical charaterizations of languages made by linguists.
Now it may seem as if idealism or instrumentalism threatens here, as if I have denied the very possibility of "getting someone wrong." You might think this if, like Dummett, you thought that only (pre-existing) shared rules can provide objective grounding for attributions of meaning. But this is false. Naturally, again, you may dispute my attribution of certain meanings to our informant's utterances; but that just means that you are not satisfied that communication has occurred, i.e., you feel that we interpreters need to continue the process of interpretation further – that I have jumped the gun. And again, just as in the other skeptical case, what you may not do is allow that communication has occurred, but that (due to the lack of previously established agreement about meaning), my attributions are somehow still suspect. It's like saying "yes, we should believe that P; but is it really true?" Compare: "yes, you two succeeded in communicating; but is that what he really meant?" In either case, to ask this is to grant something in one breath and take it back in the next (not good).
Of course people say that first thing too. And this last bit (about the parallel) is my line, not Davidson's. Davidson doesn't say much about epistemology, which leads him into some trouble by my lights, but we'll leave that for another time.