Data-driven algorithms are increasingly used by penal systems across western jurisdictions to predict risks of recidivism. This chapter draws on Foucauldian analysis of the epistemic power of discourse to demonstrate how the algorithms are operating as truth or knowledge producers through the construction of risk labels that determine degrees of penal governance and control. Some proponents emphasise the technical fairness of the algorithms, highlighting their predictive accuracy. But in its exploration of the algorithms and their design configurations as well as their structural implications, this paper unravels the distinctions between a criminal justice and a social justice perspective on algorithmic fairness. It argues that whilst the former focuses on the technical, the latter emphasises broader structural consequences. These include impositions of algorithmic risk labels that operate as self-fulfilling prophesies, triggering future criminalisation and consequently undermining the perceived legitimacy of risk assessment and prediction. Through its theorisation of these issues, the chapter expands the parameters of current scholarship on the predictive algorithms applied by penal systems.