The application of artificial intelligence as a tool to support judges’ decision-making in the process of sentencing is considered one of the most controversial contemporary developments in the realm of criminal justice. Among the array of theoretical and practical challenges raised by this development, the “input problem” occupies a central yet relatively underexamined position. This problem, in general, refers to the difficulties related to the collection, selection, and structuring of data that must be used to feed sentencing algorithms. The emergence of this challenge can be attributed, at a minimum, to two fundamental factors: first, the necessity for any recommendation-based algorithm to rely on precise, case-specific, and structured data concerning each criminal case; and second, the inherent complexity of fully and meaningfully representing the reality of the crime, the offender, and the circumstances of its commission in the form of data that are amenable to machine processing. Although the input problem has attracted the attention of researchers since the earliest stages of designing sentencing support systems, the predominant focus of the existing literature has been on its technical or functional aspects, and systematic examination of the ethical dimensions of this problem has largely been neglected. The present article, adopting a descriptive and analytical approach, seeks to fill this theoretical gap. To this end, it first demonstrates that the “input problem” is not a uniform or singular concept and has been understood in the relevant literature through different interpretations. It then analyzes the question of under what conditions and for what reasons each of these interpretations may possess independent ethical significance. Finally, by focusing on the stage of designing and deploying algorithms in criminal courts, the article offers proposals aimed at reducing the undesirable consequences arising from the complexity of the input stage.