Any technology in the nuclear industry represents variance--it may be an improvement in safety or efficiency, or it may contain some unseen defect that allows a catastrophe to happen.
But the simple possibility of harm isn't enough to swing the decision one way or the other. The application of AI methods--whether to the real-time control of plant variables, or the early detection of problems, or to the design of plants and their components--seems likely to be as beneficial as in other realms.
For example, check out the publication list of a lab active in this area. Their paper I'm most familiar with is one in which they build a fault detector paired with a fault library classifier, so that the operators can be alerted not just that something is abnormal but what fault has probably occurred. This is done in such a way that standardized plants (such as, say, the French nuclear system) can share records with each other, meaning that any plant has the experience of every plant at their fingertips.