Except AI isn't being 'sandboxed away' in any discernible way. A child being an unpleasant annoyance isn't comparable to a an AI that can't drive in either risks of failure or capability because the child is far more capable of moderating his behavior to conform to social norms than an AI.