CONCEPT
Intelligible Machines
Satya Nadella’s insistence that capability without interpretability is an incomplete engineering deliverable—the demand not just for AI that performs but for AI whose reasoning can be followed, whose failures can be diagnosed, and whose outputs can therefore be genuinely learned from.
Intelligibility is the property of being able to be understood, and Satya Nadella has argued, against the consensus of his industry, that it is a design specification rather than a research aspiration. The standard measure of progress in artificial intelligence has been capability: can the system recognize an image, translate a sentence, pass an exam? Nadella’s intervention is to add a second axis orthogonal to the first—not just what the system can do, but whether its doing it can be understood. A machine that produces correct outputs through a process no one can follow is, in his account, insufficient, and the insufficiency is not aesthetic but structural: a system that cannot account for its outputs cannot be held accountable, cannot earn trust, and cannot teach its operators what it learned from its failures. The uncomfortable diagnosis he presses is that capability and intelligibility, in the current paradigm, appear to be inversely correlated: the methods
Keep reading with YOU ON AI
Unlock the full book, 10,000+ field-guide entries, and a 1000+ thinker library. If you have a book code, register now — it takes a minute.