Semi-supervised learning (SSL) provides an effective means of leveraging\nunlabelled data to improve a model performance. Even though the domain has\nreceived a considerable amount of attention in the past years, most methods\npresent the common drawback of lacking theoretical guarantees. Our starting\npoint is to notice that the estimate of the risk that most discriminative SSL\nmethods minimise is biased, even asymptotically. This bias impedes the use of\nstandard statistical learning theory and can hurt empirical performance. We\npropose a simple way of removing the bias. Our debiasing approach is\nstraightforward to implement and applicable to most deep SSL methods. We\nprovide simple theoretical guarantees on the trustworthiness of these modified\nmethods, without having to rely on the strong assumptions on the data\ndistribution that SSL theory usually requires. In particular, we provide\ngeneralisation error bounds for the proposed methods. We evaluate debiased\nversions of different existing SSL methods, such as the Pseudo-label method and\nFixmatch, and show that debiasing can compete with classic deep SSL techniques\nin various settings by providing better calibrated models. Additionally, we\nprovide a theoretical explanation of the intuition of the popular SSL methods.\n