| Popis: |
Machines do not presently understand the world in moral terms, and therefore they act on a scene without consideration of moral implications. Asimov asserted his laws of robotics for narrative effect, but these laws also serve as a practical starting point when considering equipping machines with the ability to morally reason. Decomposing just the first clause of the first law (“A robot may not injure a human being”) reveals entailed perceptual and reasoning capabilities: What is a human being? How is a human injured? What is injurious to a human? Thus step one in setting the foundation on which intelligent machines may engage in interpersonal moral deliberation is to enable perception of these concepts of human, injury, and injuriousness. Image semantic labeling is presently indifferent to moral concerns; machines are incapable of identifying minds within a scene that may suffer and what affords or endangers those minds. Moral-Scene Assessment endows machines with a protoconscience by identifying what is morally salient in a scene, guiding interaction with those entities, and reasoning over potential harms to them. Identification of what is morally salient requires mind perception and qualification of objects of danger or affordance to those minds. With these percepts the machine is then prepared to adopt the appropriate stance in interacting with entities within a scene as it works with humans to build a shared context. With the appropriate mode of interaction selected per entity in a scene, the machine is now disposed to apply the respective ontology by which to reason over insults and injuries involving those entities. |
