The paper is interesting because CoT has been so widely demonstrated as effective. The point is that it "can" hurt performance on a subset of tasks, not that CoT doesn't work at all.
It's literally in the second line of the abstract: "While CoT has been shown to improve performance across many tasks..."