The more difficult reading is to be preferred. This principle of
codicology, or the study of manuscripts (= MSS) and the transmission of
texts, states that if you have two MSS (or families of MSS) offering different (but plausible)
readings at a certain passage, you should prefer the one which is (or seems
to be) harder to understand, as long as it does not involve bad grammar.
The idea is that a scribe may copy a text from a dirty, damaged, or sloppy
exemplar and fill in a gap in sense with something plausible. The likelihood is that
the copyist would fill in meaning with less understanding than the original
author, and he
is unlikely to invent highly difficult or ornate syntax just to fill in a
blank (copyists tended to be modestly-educated drudges).
If this happened at an early enough point in the transmission of an ancient
text, you may end up with two whole families of MSS, one exhibiting the
original reading of the archetype and the other propagating the bad reading
down to its descendents. (Some errors get in so early that they appear in
all MSS and are nearly impossible to identify unless they are ungrammatical
or make no sense at all.)
Sometimes (though less frequently) a copyist cannot understand a word or phrase
and replaces it with something sensible on the assumption that some earlier
copyist screwed the reading up. This is an emendation, and scribal emendations
are the bane of codicologists, because they go against the overwhelming majority
of entropic errors, and against the principle under discussion here.
Clever scribal emendations are often picked up by modern editors, however--it
being a mistake to think that only moderns are smart enough to get to the root
of these problems.
This principle must be applied carefully, therefore, with some knowledge of
the scribe, if indeed anything is known. I once edited a 7th-century Byzantine
astronomy text, and had to collate (systematically compare) a much later
(13th century) copy made by an accomplished (and much smarter) scientist for
his own use. Luckily, the text survives in several MSS, because my scientist's
copy was not very useful for getting at the original readings: he was constantly
emending the text, jumping over annoying repetitions, and generally adding voluminous
notes in the margin. Even worse, my scientist was a go-getter,
and had fetched other MSS of the text so as to compare readings for himself,
thus contaminating his MSS with readings from other families!