Chunking through statistical learning reportedly biases memory for features of chunked items toward each other. We examined how this bias depends on the strength and consistency of participants’ internal representations.
Across two experiments, participants completed a familiarization phase in which they uncovered hidden target objects in a grid based on a cue object. Cue–target pairs followed unknown spatial pairings with 100%, 80%, and 0% predictability in the Perfect, Partial, and Random conditions, respectively. On each trial, object colours were sampled from unique, object-specific distributions. Familiarization consisted of object–cue pairs with either Perfect and Partial (Exp1, N=20) or Perfect and Random structures (Exp2, N=18), presented in intermixed trials.Following familiarization, participants provided four estimates of each object’s typical colour, yielding two measures: deviation from the distribution mean (precision) and response variance (consistency).
Memory bias toward the paired colour emerged only in low-consistency estimates for Perfect and Random stimuli. High-consistency reports clustered around the true colour mean, suggesting that previously reported biases arise primarily from less precise representations. In contrast, consistent reports for Partial stimuli were biased away from the paired colour, whereas less consistent reports centred on the true mean, indicating that rule violations modify chunked item representations.