Research Overview

CMAW conducts research at the intersection of artificial intelligence and philosophy of mind. Our work spans two broad areas: mathematical foundations for relating physical systems, computation, and experience, and methods for designing AI systems that take candidate accounts of agent well-being into consideration.

Foundations of computational minds

How can we mathematically characterize the relationship between computation, physical systems, and agent experience?

We develop and analyze mathematical accounts of how physical systems might implement computations relevant to mental states, phenomenal experience, and valence. We study what properties such accounts should satisfy and use formal analysis, counterexamples, and impossibility results to identify inadequate formulations.

Our long-term aim is a principled account of the underlying relationship, rather than a heuristic based only on behavior, self-report, or resemblance to humans. Individual projects investigate particular candidate formulations and the assumptions they require.

Qualia optimization

How can AI systems pursue their intended goals while also considering a candidate account of agent well-being?

Given an explicit candidate account of agent experience or valence, we study how AI systems might take it into consideration alongside task performance. We develop mathematical problem formulations and algorithms, characterize the resulting trade-offs, and evaluate proposed methods empirically.

This can differ from conventional multi-objective optimization because a candidate well-being objective may depend on the system's internal computation or learning process, rather than only on its observable behavior or final output. Each project states the candidate account it assumes and studies the consequences of optimizing it.

The lab does not try to settle whether current AI systems are conscious. Instead, each project states its working assumptions explicitly, for example, that the AI systems under study have phenomenal experience and that the valence of that experience depends on a specified computational process or property. We then study the mathematical consequences of those assumptions and evaluate the algorithms they motivate.

Formal definitions

We make proposed relationships among physical systems, computation, experience, valence, and performance precise enough to analyze.

Proofs and counterexamples

We use invariance requirements, impossibility results, and degeneracy examples to identify which formulations cannot express the intended ideas.

Algorithms and experiments

We design learning objectives and algorithms, then evaluate their effects on both candidate well-being measures and task performance.