A/A testing should be used to get accurate estimates for within-sample variance. If you run an A/A/B test then you can calibrate the A/B component to be sensitive w.r.t. the tolerances of real data.

And then yeah, I'm sure a lot of successful A/B tests will get washed.