On the one hand, the current discussion on artificial intelligence and trust is characterised by what might be called “trust enthusiasm”, which conceptualises trust as an attitude that we can, and should, have in relation to AI systems when they have reached a sufficient level of maturity. On the other hand, this use of the concept of trust is viewed with great scepticism in a significant part of the philosophical literature. Two of the relevant arguments in this context are, first, that trust in AI systems is incompatible with the opacity characteristic of these systems, and, second, that to say that one can “trust” such systems amounts to a kind of category mistake. In this paper, I will argue that both the enthusiastic and the sceptical positions are problematic. Against the sceptical position, I argue that neither the opacity argument nor the category mistake argument is ultimately convincing, and that there is at least one natural use of the concept of trust – trust as an attitude of unquestioning acceptance – that can be applied to the relationship with AI systems. On the other hand, I argue that trust enthusiasm paints an overly uncritical picture of trust and tends to neglect its risks and downsides. I counter the enthusiastic picture with the principle of caveat usor, arguing that trust in AI systems should always be accompanied by epistemic vigilance.