The approach-avoidance task (AAT) has been used to help individuals avoid unhealthy foods, but results are mixed and often underwhelming. The current study explores the psychometric properties and approach-avoidance bias modifying capability of the dual-feature AAT, a novel AAT variant developed to modify approach-avoidance biases while concealing stimulus-response contingencies and requiring participants to evaluate the stimulus type. In this study, this task entailed approaching foods and avoiding objects if they were surrounded by a particular shape, while doing the opposite if they were surrounded by another shape. We contrasted it with an irrelevant-feature AAT, which involved approaching or avoiding on the basis of the shape around the stimulus, regardless of its content. In an online experiment, we used both to train participants to avoid chocolate and approach fruit. We compared training outcomes to those of equivalent sham trainings and to each other. The irrelevant-feature training was associated with more desire to eat and selection of fruit, affecting both trained and untrained stimuli; meanwhile, the dual-feature AAT only increased approach biases for trained fruit stimuli. Participants’ self-reported stimulus-response contingencies also affected desire-to-eat ratings, food selection, and bias change in both training types, principally in the active training condition. The dual-feature AAT had much higher error rates, RTs, and self-reported difficulty. The effect of perceived contingency likely reflects selective recall of specific stimulus-response associations in the task, and should be studied as a means through which to enhance the effectiveness of future AAT-based trainings. The underwhelming training outcomes of the dual-feature AAT can likely be attributed to its difficulty, as this precluded the acquisition of simple stimulus-response associations; this difficulty must be reduced before the paradigm can become effective.