The room of amnesiacs: AI individuation and reference

Three interconnected clusters of questions: 1. What sort of thing is an AI, how many of them are there, where are they and how long do they exist? 2. Is an AI the sort of thing that can have beliefs, desires and goals? 3. When we say to an LLM, “Your argument is confused”, what does that 'you' refer to? When it replies, “I'm sorry for the confused argument”, what does that 'I' refer to? In this paper, our focus is on reference and coreference: we approach 1 and 2 by first trying to answer 3. Three views are argued for. First, don't start with metaphysics or philosophy of mind-start with reference questions. Second, the practice of referring to LLMs is being engineered; it's unsettled and currently so vague that there's no fact of the matter about which of many candidates we're talking about. Third, what will make reference more determinate is a range of institutional factors: laws, regulations, authorizations and various practical arrangements. The answers don't come from the philosophy of mind or from metaphysics. They come from the institutions and practices we embed LLMs in.