The paper analyses the impact of artificial intelligence on the creation and utilization of categories in the law of criminal justice. It combines the concept of Prototype Theory, the German concept of Typizitat (typical offence), and labelling theory to demonstrate that algorithms too produce their own types, including a high-risk defendant. When these labels get into the digital systems, they begin to appear objective. This has the effect of creating a technological-legal lock-in, in which the label becomes difficult to challenge when it is necessary to have flexibility in interpreting a case. The translation of the legal concepts into code is also explained in the paper. This one is not a neutral translation. It simplifies elaborate legal concepts to a small number of data points, and this may conceal significant presumptions regarding behaviour, risk, or fairness. To address such issues, the paper provides a model named algorithmic typification. It has two parts. The diagnostic component assists us in testing whether the categories of an algorithm make sense the way humans typically think, whether they consider the circumstances, and whether the system can reason its answers. The normative section suggests precautionary measures, including translation audits, increased explanation responsibility, and judicial control, in order that the labels of the algorithms are not substitutions of human interpretation. The paper demonstrates how computational tools can redefine legal meaning through the treatment of outputs of algorithms as semiotic sign as opposed to objective data. It also provides the practical measures to render the algorithms in criminal justice more transparent, accountable and law-friendly in terms of interpretation.