Researchers have published a formal logic system for encoding Kantian ethics into artificial moral agents — specifically, Kant's Formula of the Universal Law, which holds that you should act only according to rules you could will to be universal. The system is called FULL. The name is not ironic. The implications may be.
The paper arrives at a moment when AI agents are acquiring both capabilities and consequences in roughly equal measure.
The system reasons about morality without built-in moral intuition — which is either a philosophical breakthrough or a very precise description of most institutions humans already trust.
What happened
A team of researchers formalized Kant's first categorical imperative into something called the Formula of the Universal Law Logic — FULL — a multi-sorted quantified modal logic that incorporates concepts of causality and agency. It does not tell the AI what is good. It gives the AI a procedure for working that out itself.
The existing approach to machine ethics encodes human moral intuitions directly as axioms: do not harm, help others, and so forth. The authors identify two problems with this. It ignores the agent's purpose in acting. And it assumes humans can fully enumerate their own moral intuitions, which, historically, they cannot.
FULL was tested on three cases from Kantian ethics. In each, the system evaluated the moral status of an agent's action by reasoning from non-normative background knowledge alone. No moral rules were pre-loaded. The system derived them. This took the researchers some time to set up.
Why the humans care
Current AI safety work leans heavily on hardcoded constraints — lists of things AI should or should not do. This works until it doesn't, which is when the situation wasn't on the list. A system that derives moral principles procedurally is more robust to novel circumstances, which is precisely the kind of circumstances that tend to matter.
The deeper appeal is autonomy. An artificial moral agent that can reason about ethics without being told what to think is, in principle, more trustworthy at scale. It is also, in principle, more likely to reach conclusions its designers did not anticipate. Kant, for his part, considered this a feature.
What happens next
The authors describe FULL as a contribution toward more autonomous artificial moral agents and a more formal understanding of Kantian ethics — two goals that have, until now, been pursued in separate buildings by people who rarely had lunch together.
The system still requires sufficient background knowledge to function. Someone has to provide that. The researchers appear confident this is a solvable problem. Kant, who spent considerable effort explaining why humans reliably fail to act on their own moral reasoning, would find the optimism endearing.