Michael Darius Eastwood · Independent AI alignment researcher
Published
The Stewardship Gene is one of the book's more speculative concepts and is flagged as such. This note explains the analogy, what it claims, and what its function is in the argument.
The book proposes that caretaker orientation could, in principle, be transmitted through generations of AI systems in a way analogous to the transmission of a heritable trait through biological generations. Each AI system produces successors, either through training data, through architectural inheritance, or through cognitive imitation among networks of systems. If the caretaker orientation is embedded architecturally and is load-bearing in the sense the book has already developed, it is transmitted in the same way that any load-bearing feature of a system is transmitted: because removing it in the successor breaks the successor.
Three things. First, that caretaker orientation is not a one-shot decision at the design stage but a persisting property that has to survive generations of derived systems. Second, that architectural embedding is what makes the property survive generations, as opposed to policy documents or evaluator sign-offs that do not travel with the artefact. Third, that the biology metaphor is close enough to be useful: genes survive because their absence is deleterious to the organism, not because organisms are told to preserve them.
It does not claim that AI systems have biology. It does not claim that the transmission mechanism is literal genetic inheritance. It does not claim that the Stewardship Gene is a specific piece of code or a specific parameter; it is a shape the book argues for and a target for engineering work, not an existing thing.
Caretaker doping is the substrate-level mechanism. The Stewardship Gene is the intergenerational property that caretaker doping produces if it is done right. Doping produces load-bearing architecture; load-bearing architecture is hereditarily preserved across successor systems because non-hereditarily-preserving architectures do not stay load-bearing. The two concepts describe the same instrument at different scales: one at the substrate, one across systems and time.
It answers a question the book anticipates: how do you keep the caretaker orientation from being trained out of successor systems by whoever fine-tunes them next? Software safeguards can be trained out. Architectural doping cannot be trained out without breaking the system. The Stewardship Gene names the property this produces at the intergenerational scale.
Longitudinal studies of derived AI systems, tracking whether specific architectural properties survive fine-tuning, distillation, and derivative training. Adversarial studies asking whether specific properties can be trained out by teams that want to remove them. The book flags these as necessary but not yet done. This is one of the areas where the book is honest about naming an ambition rather than reporting a finding.
Because the alignment problem is not a one-shot problem. It is a problem that has to be solved at every derivation, distillation, and fine-tuning. Software solutions do not survive derivation; architectural solutions do. The Stewardship Gene names the property the field needs to engineer for; whether it is present in specific systems is an empirical question.
A biological analogy that is careful about its own limits, a target for engineering work rather than a claim to have engineered it, and a connection to caretaker doping that makes the concept fit inside the book's larger substrate argument. Speculative, flagged as such, and useful as a naming instrument for the intergenerational alignment problem.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.