Humans can estimate their confidence in making correct decisions, but these confidence judgments are biased by their other estimations, an effect known as confidence leak. However, it remains unclear whether this effect arises automatically. Here, we address this issue by having participants make two visual decisions and give confidence ratings for one or for both decisions within each trial. Using the well-known interaction between task difficulty and response accuracy as a proxy for confidence, we found that confidence ratings for one decision were greater when the other decision was also associated with greater confidence, even when the latter was not explicitly rated. For one of the two tasks, this confidence leak also occurred when participants knew in advance that no confidence rating would be required for the other task. Our results support the idea that confidence can be automatically integrated across decisions.