Graduate Student Talk Series _ Yunlong Cao

Complex AI systems are opaque in many important epistemic and practical senses (e.g., Burrell, 2016; Creel 2020). This paper focuses on one seemingly trivial sense that contemporary computer-based AI systems are transparent: the best explanation for a complex AI systems' behaviors is always available to our expert. I argue that even though this notion of transparency might be under-satisfactory for some practical purposes (e.g., justifying AI use in decision-making), it has a philosophy of mind consequence of undercutting our justification for attributing mentality to AI based on their behaviors.

 

I show this transparency in two steps. I first argue that computer programs have a social epistemic property of transferablility just like mathematical proofs (Easwaran 2009), such that in principle, experts can derive any output of a program just based on the the program itself (together with any input or runtime environmental conditions). This is evidenced by both the software engineering practices of pair programming and open-source development, and theoretical connections such as the Church--Turing Thesis, the Curry--Howard Correspondence, and arguments for the formal verification of programs. Furthermore, such transferable programs feature in both a mechanistic explanation (Piccinini 2007) at particular output level and a distinctively mathematical explanation (Lange, 2013) at the program and algorithm level.

 

This talk will focus on the social epistemology and explanations of computer programs instead of the philosophy of mind motivation.