In experiment 1, ice demonstrate that children can learn an artificial
grammar in the standard conditions (see, for example, Reber, 1993). T
his implicit learning capacity does not develop in the examined age-pe
riod (9 to 11). Children's verbal explanations came under the three cl
assical models of artificial grammar learning: abstraction of rules, c
omparison with training exemplars, and memorization of fragments. Howe
ver, a complementary analysis shows that children's adherence to some
grammatical rules plays a major role. In experiment 2, toe use the sam
e artificial grammar and learning items as in experiment 1. But, havin
g discovered a system of imperfectly valid rules used by the children,
we constructed a new set of test items. This new set is designed for
checkmating the rule system of experiment 1 if the (new) children cont
inue to use it. The results conform to the prediction: the 10- and 11-
year-old children judged as grammatical the test items which were not,
and conversely judged grammatical items as non-grammatical. Two other
factors are taken into consideration in these experiments: the simila
rity and association (as measured by associative strength) between the
learning items and the test items. The results suggest, if these fact
ors are not confounded with grammaticality, that they have little impa
ct on the judgment of grammaticality. Furthermore, in the present rese
arch they appear rather as a by-product of children's adherence to the
rules. In the conclusion, we emphasize the questionable status - whic
h is exacerbated by experiment 2 - of the notion of an <<imperfectly v
alid>> rule. We also underline that in artificial grammar learning, th
e grammar and the learning items do not alone determine the test perfo
rmance.