People often report an item in visual working memory (VWM) as biased towards novel visual input shown during the delay, especially when people consider the novel input to be similar to the memory item. We investigated whether the magnitude of this “similarity-induced memory bias” is influenced by stimulus certainty of the memory item and/or the novel visual input. Across two experiments, participants (N=61) remembered the direction of a dot motion stimulus (memory target) over a 2.5-second delay, subsequently reporting this direction and their confidence. On 80% of trials, novel visual input in the form of a second dot motion stimulus (probe) appeared during the delay, and participants judged its similarity to the target. Stimulus certainty was manipulated by changing the motion coherence of the target and probe. As target coherence decreased, both precision and confidence of the target report decreased, indexing a decline in objective and subjective stimulus certainty. Critically, the similarity-induced memory bias was amplified when the target had low coherence and the probe had high coherence. We discovered that this amplification reflected two dissociable VWM memory errors: bias towards the probe, and replacement of the target by the probe. This implies that working memory contents can be biased or replaced by new input depending on the stimulus certainty.