Paper short abstract:

This presentation investigates the infrastructural, political, and institutional challenges posed by the conversion of judicial decisions into AI training data utilization. It draws on a a field investigation in France, including an ethnography at the supreme court of justice.

Paper long abstract:

Judicial decisions, as pivotal documents within the legal system, are intricately embedded, both institutionally and materially, in their creation space: the Courts of Justice. Often stored locally in on paper, in folders and boxes stored within judicial administrative services, these decisions are traditionnally bound to circulate inside the judicial system.

In 2016, however, the French Parliament introduced a principle of open and digital dissemination of all judicial decisions. This initiative prompted legal tech start-ups, emerging from outside the judicial sector, to develop AI models trained on these open-access court decision - in order to commercialize "predictive justice" algorithms.

Nearly seven years later, judicial decisions are still not fully available to the public in a digital format. Opening access to the approximately 4 million judicial decisions produced annually in France has proven to be a complex undertaking, presenting various technical and political challenges: identification of relevant documents, digitization, standardization, centralization processes, construction of technical infrastructures for storage and dissemination, anonymization of sensitive data...

This presentation explores the algorithmic activation of judicial case law archives, drawing from a field investigation conducted within the justice sector, which includes an ethnography of the "open data" department of the French supreme court of justice (Court of Cassation). The focus will be on:

• The diversity of actors involved in the choices that shape these digital infrastructures

• The material and technical dispositives involved in these processes

• The effects of on judicial decisions, which lose their legal specificity but acquire computational properties.