Dienes, Altmann, Kwan, & Goode (1995) for the first time used opposition logic of PDP in artificial grammar paradigm. They strictly match facial similarity among two sets of grammar and illegal strings, but found no automatic response. We speculated that facial similarity might lead to a great discrimination on two sets of grammar in learning phase, make their implicit knowledge in the measurement phase able to comply with explicit task requirements, and disguise the automatic response. Higham, Vokey, & Pritchard (2000) also used opposition logic of PDP in artificial grammar paradigm, and successfully discovered controlled and automatic response for the first time. However they only matched facial similarity between two sets of grammar but not among them and illegal strings, which lead to detect false automatic response caused by facial similarity rather than real automatic response caused by implicit knowledge, therefore their experiment was questioned by many researchers. So when the opposition logic paradigm in artificial grammar was born, it had so many defects that it couldn’t be further developed and couldn’t be widely used in artificial grammar paradigm. But if we make success of the opposition logic paradigm, it could be used to quantitatively detect controlled and automatic responses in artificial grammar learning, which might push reaserch of objective consciousness improving much. Therefore the present study attempts to solve the defects of it. The present study used single factor (experimental conditions: in-concert condition and opposition condition) design between subjects, and divided the participants into two groups: in-concert condition group and opposition condition group. The experimental materials were the same as Dienes, et al (1995)’s. But we bound negative label and grammar B together into implicit learning, and created an anti-opposition logic paradigm. In learning phase participants were asked to implicitly learn grammar A, but learn grammar B as illegal. In measurement phase participants under in-concert condition were asked to judge new strings following grammar A as legal, judge new strings similar to grammar B as illegal, and judge real chaotic illegal strings U as illegal; participants under opposition condition were asked to judge new strings following grammar A and B as legal, and judge only real chaotic illegal strings U as illegal. So under opposition condition, if participants were still unable to comply with explicit task requirements, and automatically judged new strings following grammar B as illegal (higher than baseline), then it showed that there was automatic response! So the difference between rejection rate (judgment for illegal) of grammar B under in-concert condition and under opposition condition was controlled response. Because rate of judging new strings as illegal under in-concert condition was two-thirds, and rate of judging new strings as legal under opposition condition was also two-thirds, in the two groups there was high probability bias effect which would lead any new string to be more judged as high probability judgment. The present study successfully calculated it as Pr = 0.13, and ruled it out. Then we made independent samples t test and found Bhit-Pr under in-concert condition (0.68 ± 0.18) was significantly greater than Bmiss+Pr under opposition condition (0.48 ± 0.13), t (40) = 4.46, p < 0.01, d = 1.274. Therefore we got the controlled response was 0.20. This proved that when we completely ruled out the impact of high probability judgment bias effect in statistics, the result showed that participants could indeed comply with explicit task requirements to explicitly control part of their knowledge under opposition condition. For testing automatic response, we made ANOVA for repeated measurement on A - miss - op, B - miss - op and U - hit – op under opposition condition and found there was a significant model main effect, F (2, 23) = 3.03, p < 0.01, η2 = 0.209, and a significant group main effect, F (2, 23) = 35.39, p < 0.01, η2 = 0.755. We further made a simple effect analysis for three groups, and found that Bmiss (0.35) was significantly greater than Amiss (0.18) under opposition condition, F (1, 23) = 3.91, p < 0.01, η2 = 0.145. The difference was 0.17 which was the evidence of automatic response: Although participants’ controlled response tried to agree with new strings following grammar B, but compared with the grammar A, participants would unconsciously and automatically refused more new strings following grammar B. We also did Pearson correlation between recognition discrimination DiscAB and automatic response AutoR and found no significant correlation, r = 0.32, p > 0.05, which proved that discrimination was not related with automatic response. The results showed that a negative label could be implicitly acquired by bonding with a set of implicit grammar, suggesting a new way by which explicit knowledge could be converted into automatic knowledge. Moreover, negative implicit knowledge was found to be more resistant to conscious control than a positive one (i.e., an implicit label could be easily converted from positive to negative, but difficult visa versa). Compared with the traditional opposition logic paradigm, the anti-opposition logic paradigm created in the present study was more effective in measuring pure automatic response by excluding the confounding influences of facial similarity of test items and participants’ ability to differentiate test items in different categories. And by removing the high probability bias effect in statistics, the present study detectived pure controlled response.