Or that the error distributions are equal across groups. That way you could still detect that one group is committing fraud at a higher rate, but false positive

Or that the error distributions are equal across groups. That way you could still detect that one group is committing fraud at a higher rate, but false positives/negatives occur at the same rate in each group