Research
Research Themes
Theme I: Philosophical Perspectives on Explainable and Interpretable AI
How can we understand the deliverances of AI systems?
Principal investigator: Herman Cappelen
Can AI speak, and if so, how? What sort of semantic theories–theories of meaning–are applicable to AI? What sort of metasemantic theories, theories that try to specify in virtue of what a given expression means what it does, apply to it? Based on Cappelen and Dever 2021, this project explores such questions as well as their application to such important issues as explainable AI and existential risk.
Theme II: Communicating with AI
What can recent work in philosophy of language tell us about AI communication?
Principal investigator: Rachel Sterken
Can recent (and not so recent) work on language (metasemantics, semantics, pragmatics, and the theory of communication more generally) can help us understand the linguistic and communicative abilities of artificial intelligence? Our optimistic working hypothesis is that many of the theories and models we’ve developed about language, meaning and communication can provide important perspectives on AI.
For recent and forthcoming research related to this project see the new publications on this website.
Theme III: AI: Cross-Cultural Perspectives
How can work in non-western philosophical traditions help us understand AI?
Principal investigators: Justin Tiwald and Amit Chaturvedi
To date, most if not all work on philosophy of AI is based on the western philosophical tradition and its way of understanding the core philosophical concepts, such as mind, at issue. This west-centrism risks overlooking valuable perspectives that come from other philosophical tradition. This project, bringing together experts in Chinese and Indian philosophy, as well as philosophy of AI, will show the relevance of non-western philosophy for cutting-edge work in the philosophy of AI.
Theme IV: Model Evaluation: Capabilities and Alignment
What are ML models capable of doing? How well are they aligned?
Principal investigator: Nate Sharadin
It is frequently said that ML models “can” do something, or that they are “able” (or unable) to perform some task. But what is required in order for an ML model to be “capable” of an action? This project involves empirical and conceptual work on evaluating machine learning (ML) models. In particular, it investigates effective strategies for evaluating the capabilities (and alignment) of large, general purpose, frontier ML models deployed in open-ended environments. Strategies for model evaluation are subject to several desiderata. For instance, an adequate strategy must make sense of how “benchmarks” bear on capabilities (and alignment) and how particular user interactions with models serve as evidence of capabilities (and alignment). We should also prefer an evaluation framework that makes sense of various model-specific ways of changing how models manifest capabilities (e.g., fine-tuning for LLMs). An ideal strategy for model evaluation will also be scalable (e.g. by being liable to automation), in the sense that it can be deployed to evaluate the capabilities (and alignment) of arbitrarily large (and capable) models.
Theme V: AI in the extreme
How should our theories about risk, value, and belief be revised in the face of the extreme societal changes that AI might bring?
Principal Investigator: Frank Hong
In the future, advanced AI systems may transform the world to the extreme. AI may pose extreme risks, or be the cause of extreme flourishing. They may themselves be capable of extreme levels of wellbeing. Or perhaps they would be able to sustain extremely large numbers of simulated people. Each possibility brings with it a host of philosophical questions that interface with our theories risk, value, and belief.
Questions of interest include: If advanced AI poses an existential risk, should we reallocate all our resources to reducing that risk, even if the probability of making a difference is exceedingly small? If AI can enjoy superhuman levels of well-being, would it be better if humanity were replaced by these AI? And if AI can produce extremely large simulations, how confident should we, including the AI, be that we are in a simulation, and does this have any implications on how we ought to behave?
Theme VI: Law, Ethics and Data Quality
Can we develop data design principles to build more responsible AI systems?
Principal Investigator: Boris Babic
Data collection, exploration, modeling and inference are not value neutral exercises. How, then, should we think about these data “design” choices? In this project, we try to answer this question from the perspective of a policy architect studying the infrastructure underlying modern AI systems. We aim to answer the following: What can we do if our data collection practices are biased? What does bias mean in this context? How and when should we protect privacy? Can we learn from noisy data? When should we build justifiable models?
Theme VII: AI, Music, and Creativity
How is AI destabilizing existing conceptions of creative work, the “artistic genius,” and such closely related notions as talent, skill, and virtuosity?
Principal Investigator: Rujing Stacy Huang
Has generative AI democratized music and the arts at large, “making everyone a musician/artist”? This project starts with examining the increasingly popular, techno-utopian narrative surrounding AI’s supposed democratization of art, and explores the continuing impact of AI on the nature of creative production. What is “Creative AI” and how could we situate it amidst historical and ongoing debates about artistic creativity? This project brings together scholars and practitioners working at the intersections of artificial intelligence and the creative arts (with a focus on music).