With the data deluge and rapid development of new IT systems to apply machine learning algorithms, science seems to be pushing the limits of risk assessment, and promises a heightened capacity to predict perturbations and hazardous patterns in complex systems. This short paper aims to gather lessons from the study of the development of computational tools for chemical risk assessment, to understand how computing innovations unfold, and which technological promises are actually fulfilled in matters of prediction. This is done in the spirit of addressing the questions laid out in the workshop invitation, notably “what, of all this, is realistic and unrealistic?” The paper engages with such questions as: How far can we regulate sociotechnical systems thanks to continually produced and modelled data, as claimed by specialists of computational or data sciences? How do we get to realize what is even possible and credible in new predictive practices? How and to what extent does this realization alter the development of databases, algorithms and other modeling tools applying in risk management systems? I will discuss the development and use of computational tools for the safety of chemicals from the perspective of the sociology of science and technology and regulation research. Overall, the perspective here is that of data and algorithms as knowledge systems that inform regulatory action in the area of risk and safety. In the remainder of the paper, I first outline what risk assessment is. I then touch on the digitalization of risk assessment: the rise of new computational tools and practices in this area of hybrid practice spanning predictive science and risk management. The third part is more empirical: it outlines the history of the development of tools for predicting chemical safety. The concluding part discusses the empirics and puts forward points for collective discussion, that broadly concern the path of development of digital knowledge systems and expectations that can be formulated as concern artificial-intelligence, ‘big data’-based predictions.