Abstract
This essay argues that there needs to be a shift in artificial intelligence, namely, a shift from the perspective of computation-oriented machine learning to the perspective of application-oriented machine training. With a training perspective on artificial intelligence rather than a learning perspective, issues raised by the humanities will become ever more important in the development of new artificial intelligence systems; from a training perspective, the goal is that intelligent systems act in ethical ways, solve tasks in trustworthy ways, and that system configuration be controlled by humanities researchers during a basic training phase. The central idea of Humanities-Centred AI (CHAI) is that intelligent systems, proven to be a useful support to humanities researchers in solving their humanities-research problems, will also be suitable for other application domains because the models used internally by these systems contain sufficient information about human culture. Humanities researchers can even actively shape these models, which can then be shared among agents and reused in different contexts. A prerequisite is that artificial intelligence moves away from offering so-called artificial intelligence methods for computer scientists and instead provide useful abstractions in the form of intelligent agents whose instruction or training is controlled by humanities researchers.
