The first measure of implicit knowledge used in this study, the semantic priming paradigm (WrdTask), enabled a comparison of response latencies following exposure to stereotype-related category-trait pairs with exposure to stereotype-unrelated pairs. As participants in this study were faster to respond to stereotype-consistent than -inconsistent category-trait pairs, it suggests that information about the related stereotype was activated thereby facilitating a faster response to stereotype consistent category-trait pairs. In this, it can be assumed that firstly a link was activated between the category and related information. Secondly, that as the category labels were denotative of gender stereotypes, the information activated can be assumed to be implicit knowledge of gender stereotypes. Thus the inference can be drawn that implicit knowledge of gender stereotypes was available for these participants. This finding of differential response latencies to consistent and inconsistent stereotypic information replicates those reported in other studies which also aim to activate chronically available implicit knowledge of stereotypes (Banaji & Hardin, 1996; Blair & Banaji, 1996; Dovidio et al., 1986; Foddy, 1997; Stroupe & White, 1986).
Relationships among measures of implicit knowledge revealed interesting information about the measures themselves. While responses to consistent and inconsistent category-trait parings were indices used to illustrate the availability of stereotypic knowledge, the fact that they both indicated agreement with the stereotype (participants answered yes to consistent pairings, and no to inconsistent pairings), suggested that they could both be used as indices of implicit stereotypic knowledge (Wittenbrink et al., 1997). As these two indices were consistently found to load together onto factors other than those containing explicit beliefs measures, it suggested that the construct they were measuring differed from that measured by explicit beliefs. In adopting Devine's (1989) conceptualisation of the cognitive components involved in stereotyping, this pattern of results suggests that the measures were tapping automatically activated implicit knowledge. Moreover, according to Bargh (1994), as respondents may have been aware of the stereotype-related nature of the task, but concomitantly were unable to control the activation of stereotypic knowledge, nor cease the activation once commenced, the task could be argued to be facilitating a form of automatic processing.
In addition to measurement of response latency, the WrdTask contained a measure of correct responses, a measure which indicated agreement with the stereotype. In contrast to the response latency measures, the correct response rate was found to correlate and form factors with measures of explicit beliefs. In explaining these findings it may have been the case that participants' implicit stereotypic knowledge matched their explicit beliefs about women. However, as participants possessed implicit stereotypic knowledge (as indicated by their response latencies), and their explicit beliefs or attitudes were low in sexism, then in Devine's sense (1989) these participants seemed to be inhibiting their implicit stereotypic knowledge in favour of their low prejudiced attitudes. As such the argument that the correlations between the correct response rate measure and explicit beliefs measures were the result of similar implicit and explicit beliefs is not overly compelling.
A more convincing explanation concerns the type of processing that may occur in making a decision as to the semantic relationship between the category and trait words. As participants were consciously aware of both the category labels and the traits subsequently presented (SOA = 500ms), it can be assumed that their decision to accept or reject the stereotypic relationship between the trait and the category was made with awareness of the features involved. Although the latency of their decision may have been a reflection of automatically activated available knowledge, their decision to accept or reject information on the basis of this knowledge may have entailed more controlled cognitive processing. This conclusion, however, is debatable given that participants, despite being aware of the stereotype-related nature of the task, may have been unable to control the way in which the activated stereotypic knowledge affected their decision, and this uncontrolled factor, according to Bargh (1994), fulfills a criterion of automatic processing.
This study employed another measure that claimed to tap a form of implicit knowledge, a form that is described by von Hippel et al. as implicit attributional prejudice (1997). This measure, the SentComp essentially aims to extract implicit stereotypic knowledge by assessing the tendency to explain stereotype incongruent behaviours. The greater the tendency to explain these behaviours the more implicit attributional prejudice the person is thought to have. In this study, the extent to which participants generally responded indicated that they possessed implicit attributional prejudice.
However, the SentComp was found to correlate and form factors with measures of explicit beliefs, a finding which contradicted those of von Hippel et al. (1997). As with the correct response rate measure, this result may be a reflection of the type of processing involved in explaining stereotype incongruent behaviours. In adopting Bargh's assumptions (1994), the fact that the participants were aware of the construct being measured suggests that the processing involved was automatic. Nonetheless, the gender-related aspect of the task was salient. As such, it may have been the case that the gender-related information was processed in a controlled manner which in turn may have affected the tendency to explain behaviours incongruent with the stereotype. If the tendency to explain was affected in this manner it would account for the relationship between this measure and the measures of explicit beliefs, measures which entailed controlled processing. Clearly, however, as the two versions of the task elicited differences in the tendency to explain stereotype-incongruent behaviours, the application of this measure requires further investigation.
This study also employed a measure of knowledge of gender stereotypes, the OthsVws. Earlier in this paper it was argued that while measures of implicit stereotypic knowledge typically aim to automatically activate this knowledge, it is conceivable that the content of this implicit knowledge is reflected by knowledge of the content of a stereotype, despite this knowledge being expressed in a controlled manner. As such this checklist was treated as a possible correlate of measures of implicit knowledge.
The pattern of relationships that the two subscales of this measure formed with other stereotyping measures suggested that it was, in fact, measuring similar knowledge to implicit knowledge (as indicated by its relationship to the WrdTask response latency measures). The low internal consistency of the scale and the variation in the behaviour of the two subscales can be interpreted in many ways. It may be that males have differential knowledge of male and female stereotypes, that the items in each of the subscales are differentially valenced, or that the application of this scale requires further testing. However, given that it may have been measuring similar knowledge to implicit knowledge it may be able to act as a base rate measurement when ascertaining whether certain procedures that aim to assess implicit knowledge actually do so.
Assessing these findings in light of Devine's automatic versus controlled distinction (1989), it would seem that there is some variation in automatic processing, as indicated by the two measures claiming to tap these processes (WrdTask, SentComp). However, whether or not these measures were actually measuring automatic processing remains to be seen.
With respect to Bargh's model (1994), while the measures appear to be operationalising certain elements of automaticity, it is unclear as to whether they are, in fact, measuring automatic processes as defined by Bargh. This is because he fails to specify which state of each component is required in order for a process to be considered automatic. It may be the case that by including the controlled state of an element (e.g., awareness), along with automatic states of other features (e.g., unintentionally) the measure is tapping a form of automaticity. This aspect of Bargh's thesis is unclear and as such, a revision is warranted.