Here are the drawbacks I see to using the "cur" download of Wikipedia:

1) The purported "neutral point of view" is subject to systemic bias.
The process supported by Wikipedia biases the content toward people who
are able and willing to contribute under those circumstances.  It's not
clear how to go about identifying this bias let alone its impact.

2) If a snapshot of the "cur" downloads is delayed, it will be subject
to gaming by people who submit changes to articles that have a
Kolmogorov complexity that is known to be low just to the gamers.
Wikipedia keeps prior archives of the 'cur' downloads and it is
unlikely anyone has gamed the system so far but now that we've broached
the subject there is the potential that future versions of the 'cur'
downloads will have been so-gamed.

3) The entire history of edits is about a factor of 10 larger than the
current version.  This would be a superior corpus for the purpose of
ferreting out how various points of view bias content -- and be very
relevant to epistemology, social and political sciences -- thereby
creating a superior AI capable of considering the source of various
assertions.  If this really is beyond the capacity of computers that
would be available to viable contestants then it might be necessary to
defer use of that larger corpus until more capacity or more funding for
the prize is available.

For a discussion of the downloads see:

http://en.wikipedia.org/wiki/Wikipedia:Database_download#Weekly_datab...

Given Hutter's proof my guess is we'll know when they are capable of
more compression than humans given a comparable base of knowledge.
There is always the question of how do we know how much compression
humans are capable of. I don't have that question answered.