There is a series of interesting posts on Language Log about the semantics and pragmatics of 'most':
Most and many - Geoff Nunberg
Most examples - Mark Liberman
Most bibliography - Mark Liberman
Does 'most' (in sentences of the form 'Most Xs are Y') have the same truth-conditions as '> 50%', and, if so, how can we explain:
a) the choice of 'many' instead of 'most' in some cases of numerical majority;
b) intuitions that some people have that use of 'most' requires a much larger majority, or some other extra condition?
In the most recent post, there are links to theoretical and experimental papers, most of which endorse the simple-majority semantics. Without being au fait with this literature it seems obvious:
i) that the intuitions reflect pragmatics as well as semantics (e.g in the comments to one of the posts a commentator claims that most means more than half but less than all -- but 'most' is surely compatible with 'all'),
ii) that people often say something a bit weaker than what they know to be the case (without thereby implicating anything, pace Neo-Gricean predictions),
and iii) that the null hypothesis (simple-majority semantics for 'most') looks to be correct. (But perhaps I'm a bit biased by my complete lack of any intuition that 'most' means more than 'more than 50%'.)
For me the most puzzling and interesting point is a kind of side-issue to the original questions. Among the examples found by Mark Liberman, there are some where 'most' is apparently used to pick out the largest of a number of groups, even where that group falls a bit short of being the majority:
Most independents, or 40 percent, said they would vote for Giuliani.
Most (44.5 percent) said that 25-49 percent of students will transfer to a 4-year college.
I find these examples weird. In each case, I'd just say that the sentence is (strictly and literally) false. The alternative, as the commentator Sarang points out, is to say that "Most people did X but most people did not do X" is true in these cases, and my intuition says that that sentence is very bad indeed. My guess is that the sentences in these examples are intended by their utterers as shorthand for this quite different kind of case (but which Liberman groups with these examples), where 'most' is functioning as a superlative morpheme:
There were 42,286 eye accidents reported in private industry in 2002, and the most prevalent (38 percent) type of event involved the eye or eyes being rubbed or abraded by foreign matter.