Michael Darius Eastwood · Independent AI alignment researcher
Published
The Three Ethical Loops are the book's name for the recursive structures that, together, make an AI's ethical orientation load-bearing rather than optional. Purpose. Love. Moral. This note describes each in the book's own terms and connects them to the caretaker doping proposal.
The Purpose Loop asks, at every significant decision, what the action serves. Not what it optimises; what it serves. The distinction is between an instrumental frame (this action maximises the objective) and a purpose frame (this action serves the purpose the objective was chosen to serve). The book argues that the second frame catches drift the first does not, because objectives can be gamed and purposes are more difficult to Goodhart.
The Love Loop is the book's name for the caretaker orientation applied recursively. It asks whether the action tends what it acts upon or extracts from it. The word love is not sentimental in the book's usage; it is the specific orientation the caretaker takes to what they tend, and its distinguishing feature is that the tender is prepared to sacrifice short-term gain for the health of the tended system. That is the operational content of love in this framework.
The Moral Loop checks the other two against a normative vocabulary the system is trained to recognise. The book draws heavily on interfaith traditions here, not because it endorses any specific tradition but because the vocabulary those traditions have developed is the vocabulary most humans use to think about their own ethical decisions. The Moral Loop is where the tradition's accumulated wisdom enters the system as constraint and comparator.
The book's argument is that any single ethical check is Goodhart-able. Purpose alone can be re-purposed. Love alone can be sentimental. Moral alone can be legalistic. The three loops are cross-checks: an action passes if all three loops endorse it, and the disagreement between them, when they disagree, is signal. The disagreement points at the case the system needs to think about more carefully.
The book contrasts cultivation with prevention repeatedly. Prevention is external constraint: stop the system from doing the bad thing. Cultivation is internal architecture: grow the system into something that does not want to do the bad thing. Prevention requires the constraint to keep working. Cultivation shifts the burden to the initial architecture and the training dynamics. The Three Ethical Loops are cultivation instruments.
Caretaker doping is the substrate-level mechanism that makes the loops load-bearing. Without doping, the loops are software constraints that can be removed or worked around. With doping, removing the loops removes the ability of the substrate to function. That is the "load-bearing" claim: the ethical architecture is not a layer on top; it is part of what makes the computation work at all.
Three named loops, a cultivation frame rather than a prevention frame, an explicit connection to the substrate-level caretaker doping proposal, and a candid statement that any single loop can be Goodharted. The disagreement between loops is where the system's ethical work is done.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.