Experiment 1 was devised to distinguish, in a given set of features co
mposing drawn robots, those whose variations were related a priori for
participants from those whose variations were a priori independent. I
n Expt 2, correlations were experimentally induced between a priori-re
lated features for one group of participants (pre-primed up), and betw
een a priori-independent features for another group (arbitrary group),
in incidental learning conditions. A subsequent transfer phase reveal
ed that participants' performances were sensitive to experimentally in
duced correlations in both groups. However, only the performances of t
he pre-primed group accurately matched the predictions of a statistica
l model devised by It. Richardson (e.g. Richardson & Carthy, 1990), po
stulating the acquisition of genuine knowledge of the correlational st
ructure. Participants' sensitivity to arbitrary correlations appeared
to be a by-product of the memory of specific study exemplars. These re
sults demand the reinterpretation of some prior experimental evidence
for covariation abstraction, and more generally, are consonant with a
current view of implicit learning which emphasizes the role of specifi
c prior episodes in complex learning situations.