Computer Simulations as Scientific Instruments

Abstract

in partial fulfillment of the requirements for the degree of Doctor of Philosophy.

What kind of trust does AI deserve, if any?

“…In part, the growth of this multidisciplinary literature has been fueled by institutional incentives and foci such as the European Commission's "Ethical Guidelines for Trustworthy AI" [28,79,81] as well as by the industry's drive towards compliance and public relations. 3 I thank one of the anonymous reviewers for this paper for requesting a more explicit definition of the term 'epistemic', as well as Nico Formanek who invited me to do the same for the term 'AI'. 1 An easy way to understand the importance of these distinct capacities is by placing them in contrast to the kinds of artifacts that enhance other kinds of capacities.…”

“…This is particularly the case for transparency and explainability [87]. While it may be intuitive to couple trust with transparent or explainable systems, their relationship is not immediately obvious and has been shown to require substantial philosophical analysis to make its case (see [2,3], as well as Ferrario and Loi [27] for limitations surrounding the relationship between transparency, explainability and trust).…”

“…For example, the European Commission's ethical guidelines places emphasis on ensuring proper incentives and safeguards are in place mainly in service of the ease of adoption [45]. Alvarado [2,3] frames the gist of these attitudes towards trust in AI in the following way: "Much of the industry's efforts in understanding trust in AI technologies has been centered on sociological concerns and observations on the psychology of trust. That is, the inquiry centers not on how to make more trustworthy artifacts, but rather on how to make people trust artifacts more easily.…”

“…While deep neural networks and other machine learning techniques will certainly have important differences amongst them that will need to be fleshed out at some point or another, in this paper I offer the first step to understand them all as subjects of a particular kind of trust, namely a trust that treats these technologies as information providers or knowledge conveyers in virtue of their prowess in data analytics. 3 Finally, as I will also argue, it follows that, even after it is established that epistemic trust is the only legitimate kind of trust to allocate to epistemic technologies, it does not mean that AI can, in fact, be trusted just yet. Whether or not AI is indeed worthy of such trust is a different question and one that may yet to be answered and hence may yet be answered in the negative.…”

“…Whether or not AI is indeed worthy of such trust is a different question and one that may yet to be answered and hence may yet be answered in the negative. As Alvarado [2,3] and Symons and Alvarado [82,83] point out, there are good reasons to think that trust and reliance on novel scientific artifacts is a non-trivial enterprise with distinct considerations even to those in other established fields of inquiry such as science in general. That is, even if we have reasons to trust a method, a practice, some individuals or an institution, these reasons and this trust are not automatically transferred to newly introduced artifacts in that same practice, used or developed by that same individual or that institution.…”