Conclusion
We have argued, or perhaps more accurately, have tried to indicate, that one source of discomfort or fear about AI is the feeling of uncanniness it can provoke. The uncanniness is a result of the presence of a kind of animation we are used to associating with human embodiment in a non-organic material form. This is not only a different form, but embodies animation in ways with which we are unfamiliar in our previous experience in animated beings. While we have come to understand that humans are not the only animals with the ability to use language, the ability of a machine to use language in a similar way and at a similar level, is uncanny. Furthermore, the animation of AI machines is embodied in ways unlike the embodiment of animation with which we are familiar. For example, the GPT2 could be accessed and immediately “present” in any computer, and downloaded from computer to computer, whereas human language use is tied to specific minds and bodies in which it develops over time.
We have tried to make use of AI in our presentation, as well as talking about it. We deployed machine-learning methods as an experiment in using digital resources that are commonly regarded as simple forms of AI technology. Specifically, we used topic-modeling to analyze selected twentieth-century philosophical articles to test whether mind-body dualism is as stable and persistent an idea as we hypothesize. The models we used are rudimentary, and they serve more as illustrative instruments than conclusive ones. We realize that this is not really an adequate way of demonstrating pervasive cultural mind/body assumptions; and it does not demonstrate that blurring mind/body distinctions feels uncanny to people. Our topic model is, at best, indicative of the kinds of machine tools that scholars in the humanities can use in their research.
Second, we hope to demonstrate the uncanniness of AI by using the Generative Pre-Training Language Model developed by OpenAi (GPT2). GPT2 has been described as a form of language processing capable of finishing your sentences for you. Our hope here was to demonstrate, rather than just write about, the experience of interacting with a machine that can produce the kinds of language that we, as scholars, have been used to associating only with ourselves as human scholars.
Possible future continuations of our research might include better ways of getting at pervasive cultural assumptions about the association of specific kinds of animation with specific kinds of bodies. Ideas include include scraping Reddit chat rooms or Wikipedia or Twitter other large pools of data to ascertain prevailing attitudes towards minds and embodiedness across a much wider swath of discussions than academic philosophy journals. We also think we could get at historical instances of uncanniness in various ways. For example, performing sentiment analysis on the literature of Spinoza’s contemporaries might show that part of the negative reaction to Spinoza can be characterized as the uncanniness of his arguments about monistic substance. In addition, we are working on an NLP model that is trained more specifically to respond to and produce scholarly essays than the Talktotransformer model we borrow here.