AI systems act more and more autonomously; they take decisions with a longer and longer horizon, and are affecting more and more of our lives, in good and less good ways.
To the limit, they are even subtly influencing and shifting our ethical and aesthetic values (a phenomenon I call the value lock), eventually reshaping us as individuals, communities, societies. Maybe that’s what we want: be more productive, less challenged, more cocooned, ok with our values to shift like they have so many times in the brief history of humanity. But are we thinking through what we really want?
We can imagine that our values are enshrined in our constitutions, democratic systems, and social lives, so that it takes deliberate effort to change them.
Yet, the normative and regulatory frameworks designed to protect these commitments are fragile, static, mostly built for a different era. Yes, I mean international law, but I also mean a liability system built for static products. We are developing and increasingly interacting with essentially dynamic agents. There is a conceptual gap here.
CERNAI exists to do the foundational work that this gap demands: provide formal models of normative reasoning, philosophical analyses of agency, risk, and responsibility, and direct engagement with the regulatory frameworks that govern AI in practice to keep control of our lives. CERNAI is composed of senior and junior researchers from computer science, law, philosophy; together associated professional and policy fellows.
The contributions here will be short, accessible, and substantive: deeper than policy commentary, lighter than a journal article. They will target anyone working in AI governance, safety, regulation, or applied ethics who wants to think more carefully about the normative foundations of what they do. Posts will be written by CERNAI members and associates, and occasionally by invited contributors.
The first contributions reflect the range of problems CERNAI is currently working on: governance, applied ethics, and technical evaluation. The first will address the particular challenges that evaluating and mitigating agentic risks pose on top of standard risk management systems, and provide a blueprint for agentic risk management. The second will offer an empirical comparison of legal competence across open-source and proprietary models, and the evaluative framework required to judge it. The third will explore the prospects and limits of AI-assisted moral reasoning.
If any of these issues sound interesting or important to you, dear readers, subscribe to the mailing list, get in touch if you work on related problems, and do share the blog with colleagues who might find it useful!
Federico Faroldi