This article is from the sci.lang FAQ, by Michael Covington (mcovingt@ai.uga.edu) and Mark Rosenfelder (markrose@zompist.com) with numerous contributions by others.
Quite a lot, if by "prehistoric" you'll settle for maybe 2000 years
before the development of writing. (Language is many thousands of years
older than that.)
Languages of the past can be recovered by comparative reconstruction
from their descendants. The comparative method relies mainly on
pronunciation, which changes very slowly and in highly systematic
ways. If you apply it to French, Spanish, and Italian, you
reconstruct late colloquial Latin with a high degree of accuracy;
this and similar tests show us that the method works.
Also, if you use the comparative method on unrelated languages,
you get nothing. So comparative reconstruction is a test of whether
languages are related (to a discernible degree).
The ancient languages Latin, Greek, Sanskrit, and several others form
a group known as "Indo-European." Comparative reconstruction from
them gives a language called Proto-Indo-European which was spoken
around 2500 B.C. Many Indo-European words can be reconstructed with
considerable confidence (e.g., *ekwos 'horse'). The grammar was
similar to Homeric Greek or Vedic Sanskrit. Similar reconstructions are
available for some other language families, though none has been as
thoroughly reconstructed as Indo-European.
 
Continue to: