What Could a Robot Know About? The Discovery of the Mind in Language

Date: March 17th, 2021
Duration: 55m:9s

The development of deep neural networks (DNNs) has significantly advanced artificial intelligence, with machines now able to carry out complex tasks that, in some cases, appear to exceed human ability. However, the underlying operation of DNNs is opaque (not readily interpretable), and—despite being generally reliable--prone to somewhat unpredictable and possibly serious errors. This has resulted in significant efforts to develop systems that can provide meaningful explanations of their functioning, so-called “explainable AI”. Such systems would account for their behavior like human beings do, i.e., in terms of reasons involving beliefs/knowledge, attitudes and/or desires. Discourse analysis work reveals profound challenges facing these efforts. Ascriptions of what others know/believe, think, feel or want have a clear meaning only in relation to certain assumptions about how people can be expected to behave (and not behave). These assumptions are prominently reflected in judgments about whether a person’s behavior reflects genuine understanding, or merely rote training. A number of recent cases of DNN behavior suggest that the assumptions in question do not hold for these systems.