Frequently asked questions

No. CMAW research treats possible AI experience as a conditional assumption rather than a conclusion. Prof. Thomas regards the question of whether present or future AI systems have phenomenal experience as genuinely open and has no current leaning in either direction. Individual researchers may hold their own views; the research program does not require a shared answer.

Projects state their working assumptions explicitly, including when they assume that the AI systems under study have phenomenal consciousness, and ask what mathematical or algorithmic consequences follow. The conclusion concerns that conditional problem, not an assessment of whether a particular current or future system is conscious.

Each of us has first-person access to our own experience. We do not have comparable access to anyone else's. We can observe another system's words, behavior, brain activity, or physical construction, but none of these observations shows that phenomenal experience is present. Moving from such observations to a claim about experience requires additional philosophical assumptions. This is the classic problem of other minds.

That limitation applies to humans, non-human animals, and machines. Treating biological similarity as decisive would risk what philosophers call substance chauvinism: assuming that brains can support mental states or experience but functionally comparable systems made from different materials cannot, simply because of what they are made of. Functionalism and multiple realizability provide reasons to resist that restriction.

Even with this uncertainty, we routinely make decisions that take the possible experiences of others seriously. We do this for other people, including people whose bodies or behavior differ from our own, and many people extend moral consideration to non-human animals. In ethics, an entity whose welfare matters for its own sake is often described as having moral status, or as being a moral patient. CMAW does not decide which systems qualify.

Conditional technical work can still identify which proposed mathematical measures of an AI agent's experiential valence fail, which formalizations remain viable, and how interventions intended to improve such a measure trade off against task performance, computational cost, or other practical constraints. An intervention that substantially improves a plausible well-being measure with negligible cost presents a very different practical choice for system designers and policymakers from one that requires major losses in performance or substantial additional resources. These results can inform later decisions about whether and how possible AI well-being should be considered, without claiming to settle whether any particular system is conscious.

Yes. There is growing excitement around research on artificial minds and possible AI well-being, with activity across AI companies, nonprofit organizations, and academic centers. Anthropic has established a model welfare research program; Eleos AI Research studies possible AI sentience, well-being, and moral patienthood; and the NYU Center for Mind, Ethics, and Policy conducts and supports research on biological and artificial minds, including consciousness, welfare, and moral status.

Researchers from this broader community are also meeting across disciplinary boundaries. UMass hosted the Philosophy and Reinforcement Learning Symposium in May 2026, bringing together researchers in reinforcement learning, philosophy of mind, and cognitive science. A separate Reinforcement Learning and Philosophy Workshop was held in Montreal immediately after RLC 2026. It was organized by John D. Martin of the Openmind Research Institute and the University of Alberta, Julia Haas of Google DeepMind, and Nishanth Anand of ExperienceFlow and Mila.

These groups and researchers do not share a single theory, method, or conclusion. That diversity is useful. It shows that questions about artificial minds are becoming a genuine interdisciplinary research area, with room for mathematical theory, machine learning, philosophy, cognitive science, empirical work, ethics, and policy. Bringing these perspectives together helps expose hidden assumptions, compare competing approaches, and identify questions that no one discipline would formulate as clearly on its own.

Yes. CMAW draws assumptions, concepts, and distinctions from philosophy of mind, and its formal results may in turn clarify the definitions and implications of computational theories of mind. Its primary technical objects are computations and learning systems: we model how physical systems implement computations, analyze the mathematical properties of those models, and develop and empirically evaluate AI algorithms. Its primary institutional home is computer science because the work centers on formal models of computation and the design and analysis of algorithms.

Projects may also draw on neuroscience, cognitive science, psychology, ethics, and other fields when those areas provide relevant constraints or candidate mechanisms.

Different projects adopt different explicit working assumptions, commonly drawing from physicalism, functionalism, and computational theories of mind. These premises are stated so readers can see exactly which conclusions depend on them.

No. Qualia optimization studies candidate well-being objectives alongside task performance, computational cost, and other constraints. It asks where trade-offs occur, whether some improvements are nearly free, and whether apparent improvements are degenerate or depend only on an arbitrary representation.

This technical problem does not by itself prescribe how much weight society should place on possible AI well-being.

No. CMAW does not endorse a rule for trading off human well-being against the possible well-being of AI systems. That is a broader ethical and social question for work beyond this lab, including ethics, social science, policy, and public deliberation.

CMAW can supply technical facts that inform those decisions. We can show that some proposed measures, such as raw reward magnitude or cumulative temporal-difference-error magnitude, have serious formal flaws; identify alternatives that remain plausible under stated assumptions; and measure how interventions affect task performance, cost, and candidate agent valence. A change that improves a candidate well-being objective with essentially no performance or cost penalty presents a very different practical choice from one that causes substantial degradation. Establishing these facts does not settle the normative question, but it can make later decisions by practitioners and policymakers better informed.

Individuals and organizations can support CMAW through research grants or gifts to the University of Massachusetts. Interested parties should contact the lab director, Prof. Thomas, at pthomas@umass.edu.