In contemporary societies, the processes of transindividuation by which knowledges are transformed into cycles and rhythms of metastability have been dramatically short-circuited. In turn, this has provoked the spiritual misery and pseudo-fabulations so prevalent all around us, including our educational contexts. For Stiegler, this is nothing short of a noetic reticulation that deprives us from ways of thinking ourselves beyond or outside of our digital experience. But digitality has not only intensified the commodification of knowledges ("savoirs"), it has also rendered even knowledge production automated, recursive and probabilistic, the uncritical implementation of ChatGPT being a prime example. What this means is that knowledge and "knowledge production" have been subsumed under the rubric of "recursive optimization" for "predictive performance." To understand this transformation, I discuss the implications of the widespread use of Bayesian statistics in machine learning. My argument is that we need to develop new speculative tools aimed at what is known as priors in Bayesian models, which is to say the probability of an occurrence before the collection of new data. What this means for education is that we need to address not only the effects of automation, but also the very conditions that give rise to these.