Flaky tests that non-deterministically pass or fail without any code changes constitute an impediment to regression testing. To understand when and how flaky tests can be detected most efficiently, we analyzed the commit histories of known flaky tests. We find that 75% of flaky tests are flaky when added, indicating substantial value for developers to run detectors specifically on newly added tests. The percentage of flaky tests that can be detected early increases to 85% when detectors are run on both newly added and directly modified tests.