Herman Cappelen and Ernie Lepore, "On an Alleged Connection Between Indirect Speech and the Theory of Meaning"

The mistaken assumption (MA) shared by almost all semantics is that indirect reports of what is said are relevant to the semantic content of sentences. Reports of what is said are deeply context sensitive (both to features of the context of utterance and to features of the reporting context), whereas semantic content is supposed to reflect context-invariant features of linguistic activity.

Donald Davidson, "On Saying That"

Davidson points out the traditional problems caused by indirect reports of speech. On one hand, we want to retain our "semantic innocence" with regard to the sentences coming after "that" clauses--we want them to mean, and refer to the things they normally mean and refer to. They shouldn't, pace Frege, refer to their ordinary senses, for example. On the other hand, the indirect report of speech does not have the substitutional/inferential properties of the unembedded sentence itself. It won't stay true when co-referential expressions are intersubstituted, for example. So what's going on? How do we reconcile these two important, undeniable features of indirect reports of speech? Davidson's suggestion is ingenious. He suggests that indirect reports should be understood as compoundings of two distinct sentences. So when you report that "Galileo said that the Earth moves", you are actually uttering two sentences: "Galileo said that" and "The Earth moves". The second sentence retains all of its normal features. The first sentence is true just in case the two-place predicate "said" is true of Galileo and that, where "that" is a demonstrative that refers to the following sentence. The "said" predicate is a primitive predicate, which picks out the "samesaying" relation. There is no semantic relation that makes my saying of "The Earth moves" and Galileo's correspond (not synonomy, e.g.). It is the primitive samesaying relation that does so. So it is a mistake to think that the compound sentence will reflect any of the properties of the unembedded sentence by itself--the first part ("Galileo said that") will have a truth value that varies with changes in the demonstrated object--the second sentence.