Simulated data is provably more vulnerable to adversarial attacks. It only gives the illusion of more data (quite literally), and should not be used for mission-critical debiasing. A NN trained on mostly simulated minority faces and mostly real non-minority faces is a nightmare, the worst of both worlds: a plausible deniability that the algorithm is unbiased, and extreme susceptibility to adversarial attacks that take advantage of this illusory unbiasedness.