It's frankly impossible that they have a training set without any biases, though I'm sure they worked to eliminate the ones they could think of (which itself would have bias).
You can make the distribution different by eg duplicating some of the data a lot, which you might want to do to improve the actual purpose of the model (translation). Any other purposes (having opinions on gender) is just a coincidence and not being optimized for/regression tested.
The alternative is they are carefully curating a training set (google historically unwilling to do anything manually) or writing one themselves???