Research-Test

The value-alignment problem for AI asks how we can ensure that the ‘values’ (objective functions) of artificial systems align with humanity’s values. One component of this problem is technical (how do we encode values or principles in AI systems?), and one component is normative (what values or principles are the ‘correct’ ones to encode in AI systems?).

One of my main research projects is an interdisciplinary treatment of value-alignment problems for AI systems. Although an intuitive understanding of value alignment is superficially useful, it should be clear that it raises more questions than answers:

• Are the ‘values of humanity’ individual values? Aggregate values?
• Whose values are they? How should they be determined?
• How can a conception of value alignment deal with the variation of values across culture or time?

My approach to value alignment is differentiated from the extant literature in machine learning insofar as, in my analysis, value-alignment problems arise from the dynamics of multi-agent interactions.

Thus, rather than focusing on which values are the right ones and how to implement them, I discuss the socio-dynamic contexts that give rise to value-alignment problems in the first place. This treatment provides a conceptual basis for understanding what value-alignment problems are (and how they are generated) in addition to shedding light on similar features between value-alignment problems for AI systems and other dynamic interactions (e.g., between human agents).

One key consequence of this approach is that, outside of the intentional misuse of AI systems by bad-faith actors, every problem arising in machine ethics—bias and fairness; transparency, explainability, and opacity; control problems, etc.—can be cashed out in terms of value alignment.

Understanding value-alignment problems in terms of their structural features underscores that we will not find solutions in the manipulation or re-configuration of training data or alternative algorithms.

In recent and ongoing work, I argue that linguistic communication is necessary for robust value alignment. Therefore, understanding the fundamental principles involved in the (biological or cultural) evolution of effective communication may lead to innovative communication methods for AI systems. The importance of linguistic communication for coordinating social values and norms in complex social systems is often taken for granted in contemporary research on value alignment for AI. However, there are important evolutionary connections between social norms and linguistic communication.

One goal of this research is to demonstrate the importance of understanding the biological origins of language as a path to achieving moral behaviour in AI systems.

Along with a multi-disciplinary research team I have been reviewing the literature on artificial moral agency. We seek to provide a systematic account of the most promising approaches to developing artificial moral agents, addressing normative pluralism, conflicts, and codification. A central research question outlines the key issues and most promising implementation approaches for high- and low-level agents. Additionally, we explore the main benchmarking approaches for assessing the level of functionality, accuracy, or moral competence of artificial moral agents. My recent single- and co-authored research in this area has sought to critically examine the very idea of benchmarking ethics for AI systems. I explore why current approaches in computer science fail in light of category mistakes involved in using moral dilemmas from philosophy as benchmarks for whether an AI system acts morally, in addition to discussing metaethical problems for benchmarking ethics for AI.

My work on the evolutionary origins of language stems from my dissertation. This work aims to address the following questions:

• What are the salient differences between the simple signalling systems that are ubiquitous in nature and the linguistic communication systems that are unique to humans?
• Which of these salient features of natural language provides an empirically plausible target for explaining how linguistic communication systems may have evolved out of simpler systems of communication?

Many researchers think that if we could explain how some distinctive feature(s) of language evolved, we would have taken great strides in bridging the evolutionary gap between simple communication and natural language. The most common feature of natural language appealed to as a gap-bridging explanatory target is compositionality (and related features like hierarchy and recursion). I argue that the emphasis on compositional syntax in language-origins research is misguided. I examine the inherent asymmetry between the benefits of compositional syntax for senders and receivers in a signalling-game context. I further discuss the binary nature of compositionality, which precludes a gradualist explanation for how it evolved. Using comparative methods from evolutionary biology, I show that there is no empirical evidence for any relevant proto-compositional precursors in nature. This body of research suggests that it is a mistake to assume that since compositional syntax provides a crucial difference between language and simple communication, research on language origins must, therefore, centre on the evolution of compositional syntax itself.

Instead, I propose that reflexivity—the ability to use language to talk about language—provides a plausible alternative explanatory target for language-origins research. Communication is a unique evolved mechanism to the extent that it can overtly influence the evolution of future communication. Once individuals learn to communicate, they may use those abilities to influence future communicative behaviour, leading to a positive feedback loop. I have demonstrated, via formal models, that reflexivity gives rise to rich compositional structures from which genuinely compositional syntax can emerge; but it emerges as a byproduct rather than an explicit target of evolutionary pressures. I further argue that reflexivity does not succumb to the problems that compositionality faces: role asymmetries are accounted for by the underlying mechanisms that give rise to reflexive communication systems; there exists empirical evidence of plausible precursors to reflexivity in nature; the precursors of reflexivity are genuinely graded. My research in this area provides initial evidence that reflexivity is a fruitful direction for explanations of language origins.

My work in social dynamics has also examined general (non-linguistic) social phenomena. This research stems primarily from work done under a grant from the National Science Foundation (USA) on Social Dynamics and Diversity in Epistemic Communities (PI: Cailin O’Connor, UC Irvine). We use formal models and simulation results to precisify arguments about various social phenomena. For example, my research on discrimination shows how small degrees of power give rise to radically inequitable distributions of resources between perceptibly distinct (but otherwise effectively identical) populations of individuals. This result has consequences for the viability of accounts of distributive justice. My research in social epistemology shows how false beliefs—concerning, for example, scientific information or ‘fake news’ items—spread and persist in a network of individuals, even when an explicit retraction is issued [18]. My future work in this area will provide novel models for analysing the social dynamics of epistemic accuracy, diversity, and inequity. This research programme is unified by the aim of understanding the role of social structure and interaction in epistemology and justice. With my co-author, Aydin Mohseni, I have also used stochastic models from game theory to analyse the cooperative challenge for AI Ethics guidelines.

My work attempts to clarify the role of the philosophy of autism by distinguishing it from autistic philosophy. The philosophy of autism is work done by neurotypical philosophers who explore philosophical questions about autism spectrum disorder, while autistic philosophy is the work done by autistic philosophers themselves. For example, I am actively developing an account of the philosophy of autism that relies on an autistic philosophy, an account of autism informed by the lived experience of autism. The philosophy of autism also connects and overlaps with social, ethical, and political issues regarding autism and autistic individuals.

Autistic philosophy is philosophical work done by autistic individuals, who often approach philosophical problems from a distinct phenomenological and experiential perspective. The goal of autistic philosophy is to develop knowledge about autism from within the autistic community, rather than imposing neurotypical frameworks on autistic experience. This work often involves first-person testimony, critique of dominant narratives, and the development of alternative epistemologies and ontologies related to autism.