Don't fear the unlabelled: safe semi-supervised learning via simple debiasing