In two experiments we employed calibration methods to investigate the reali
sm of participants' confidence ratings of their own classification performa
nce based on knowledge acquired after training on an artificial grammar. In
Experiment 1 participants showed good realism (but overconfidence) for gra
mmatical strings but very poor realism for non-grammatical strings. Method
of training (string repetition in writing or mere exposure) did not affect
the realism. Furthermore, the participants underestimated their overall per
formance. In Experiment 2, using a more complex grammar and controlling for
two types of associative chunk-strength, participants showed good realism
(but still overconfidence) for both letter and symbol strings, irrespective
of grammaticality. Together, these experiments show that implicit learning
can give rise to knowledge products that are associated with fairly realis
tic meta-knowledge. It is argued that both the zero-correlation criterion a
nd the guessing criterion are misplaced when used to define implicit knowle
dge; two reasons being that confidence judgements may be affected both by i
mplicit knowledge and by inferences.