Abstract
While machine learning techniques have been used to model categorization/decision making tasks that are beyond the capabilities of traditional AI, these new models are typically uninterpretable, i.e., the reasons for their decisions are not clear. Some have argued that, in developing machines that can report the reasons for their decisions, developers should take, as a guide, human explanations for behavior, which make reference to mental states (e.g., knowledge/belief). This proposal is correct, but unattainable given certain characteristics of current AI. To explain, this article draws on recent discourse-analytic research showing that ascriptions of knowledge/belief presume behavioral performances to instantiate particular sorts of broader dispositions. This is reflected by the possibility of ascribing knowledge/belief to an agent on the basis that there can be no other explanation for their observed behavior. The behavior of AI trained through machine learning is unpredictable in ways that precludes such certainty. Consequently, while it is certainly possible to program machines to report mental states of knowledge/belief to account for their decisions, the failure of current AI to engage in typically human forms of life means that such ascribed mental states are inevitably meaningless.