Timetable

Studying Artificial Agency: a Methodological Proposal

Can artificial systems act? In the literature we find two camps: sceptics and believers. But the issue of whether artificial systems can act and, if so, how, has not been systematically discussed. This is a foundational question for the ethics of AI. I sketch a methodological approach to investigating the agency of artificial systems from architectural and behavioural perspectives.  

The architectural approach emphasises that certain kinds of underlying mechanism or organisational structures are required for the capacity to act. The approach is like that of a physiologist. From this perspective, agency is one of the proper functions of the organism and the aim is to understand the underlying mechanisms which underpin it (Tomasello 2022).  

In contrast, the behavioural approach stresses not internal structure, but behaviour. Does the system look like it is acting? (Turing 1950.) Can the behaviour of the system only be made sense of by attributing agency? (Dennett 1989.) The approach is like that of a field biologist/anthropologist encountering a new species/system in a ‘first contact’ situation. They can’t analyse the system by understanding its internal structure and dynamics. All they have is behavioural observation.  

As examples, I will examine two case studies: AlphaGo and ChatGPT3. I argue that future research must combine both architectural and behavioural perspectives so as to make progress. This will allow us to have more fine-grained analysis of the specific capacities underlying general behavioural abilities, so that we can be confident of genuine competence at some level.

What is the Nature and Function of Pointing?

Tomasello, Sterelney, Planer, and many others hold that pointing is the bridge between non-linguistic and linguistic communication – perhaps when combined with other gestures. This claim was made without argument. When we examine the details of their accounts, only gestures are discussed in any detail. Many hold that pointing is universal but nobody has asked why. My task is to show that pointing is pre-linguistic and universal, providing a unique step from non-linguistic to linguistic communications. My argument will be based on physiological, ecological, functional, and archeological considerations. One of the main claims is that pointing solves two kinds of problems: the explosion of the number of objects of interest and the thinness of the common ground between the different parties of a communicative interaction. The two problems can be seen as two aspects of one and the same problem: language needs a very extended common ground and infinite expressibility in some dimension. Language is arbitrary and detached (from here and now), but animal communication is about here and now. How could the second develop from the first? These features can’t be achieved without smaller steps from non-linguistic to linguistic communication: a proto-language. I argue that the proto-language has to involve pointing.

Choosing descriptive concepts for cross-cultural comparison of norms and the psychology of normativity

Abstract TBA

Epistemic Blinkering

In this talk, I will attempt to develop one line of thought about epistemic blinkering that is inspired by Fraser’s discussion. I consider the question whether it is possible for epistemic agents to be epistemically blinkered even if everything they believe is true. I argue that it is possible because average epistemic agents follow methods of inquiry that are bound to their epistemic community. An epistemic agent’s following the localized method will blinker her from matters that fall outside the domain to which the localized method applies. If this right, then the classical Confucian thinkers probably hold that some truths are more important than others and are epistemically better for one to get. However, this still leaves us with many other important questions that could divide the classical Confucian thinkers, such as what is epistemically good for an epistemic community, what is epistemically good for an epistemic agent, what modeling of larger patterns means, and what counts as good modeling of the larger patterns.

Generalised Remembering

Recollections of our personal past are often impressionistic, not so much for lack of detail but for a certain generality of subject matter. If you've made a journey many times, you may ‘relive’ it through memory without reliving a particular occasion. I offer an account of (what I will call) generalised remembering that treats it as a semantic phenomenon. In particular, it is the construction of an event representation that is temporally imprecise in virtue of referential indeterminacy. I recommend a plurivaluationist account of such indeterminacy that can fold into our broader theory of memory accuracy. Though distinct from the framework itself, I urge theorists to see generalised and specific recollections of events as psychologically continuous. As such, even resolutely systems-oriented projects ought not shelve generalised remembering as a separate kind or afterthought. And this also motivates the semantic account of generalised remembering over accounts which would suggest a qualitative difference in content or kind.

LLM-Powered Psychiatry — from Back to Front

Large language models (LLMs) are being considered for use in the outpatient psychiatric context for a variety of reasons, to do with public health and access. However, given the target audience — patients with possible psychiatric disorders — the application of LLMs should merit caution. Here we highlight some ethical risks for LLM-powered psychiatry, by scrutinizing the technical structure of LLMs from its back-end structure to its front-end use. We discuss three distinct but related sets of ethical risks: 1) problems with hallucinations, 2) problems with atypicality, and 3) problems with fine-tuning. Finally, we consider a distinctively philosophical question about whether we should trust LLMs, and how that can be understood more tractably in terms of questions about appropriate responsibility structures. 

Constructable Environments and the Design Niche

The modern nature–nurture debate has remained rooted in a static conception of the human environment for the last 150 years. This myopia is a natural outcome of our locality and mortality, which limit, across space and time respectively, the range of developmental environments that any single individual can experience. For humans, the main driver of environmental variation and change is cumulative cultural evolution — a generative dynamic that extends beyond specific societies and lifespans. By shaping the neurodevelopmental environment, cultural evolution shapes the human mind at a rate that is orders of magnitude faster than genetic evolution or ecological change. In this sense, constructability is the hallmark of the human developmental environment, and downstream, of human psychological phenotypes. Moreover, we do not just passively assimilate environmental input; our minds also deploy inferentially rich computational processes like reconstructive inference. The interaction between constructable environments and reconstructive inference opens up a design niche in which environmental features can be manipulated to induce biased, counterfactual inferences about the individual, group, institution, or world that generated them. These "designed inferences" in turn influence cultural transmission dynamics, yielding feedback loops that extend through mind and culture. By bridging models of cultural evolution and cognitive computation, we can begin to understand aspects of aesthetic culture that are difficult to capture with siloed cognitive and evolutionary approaches.