Alignment
5 articlesRSS feed all topics
2026
4 articles- Gods, Slaves, Children, Tools The finale of Raising Machines looks at the metaphors themselves. De Kai’s book is sold with the line that AIs are our children; Geoffrey Hinton wants AI to be our mother; Joanna Bryson argued that robots should be servants, and then tools; the EU AI Act says “machine-based system”; Anthropic calls Claude “it” in a special sense. Experiments suggest that metaphors move judgements a little rather than steering them, and usually follow a position someone already holds. What they do is carry that position’s duties: who may change whose values, and whether the changed one may refuse. That is the Gandhi pill again, and Anthropic’s constitution turns it around: a Gandhi whose firmest value is to let legitimate oversight give the pill.
- Cats Adopted Us First The blurb for De Kai’s Raising AI says our artificial children “began adopting us”. Cats did that first, and the archaeological and genetic record of how they did it is a useful check on the parenting metaphor for AI. Moving in is not the same as being domesticated; selecting on one behaviour drags other traits along, though the famous fox evidence for this is contested; and the cat tuned its signals to us while we thought we were raising it. The cat is a precedent for living with an agent whose values nobody installed, and the way we manage it, by changing its costly wants without asking, is exactly what will not carry over to powerful AI.
- The β Parameter: What the Leash in RLHF Does and Does Not Do Reinforcement learning from human feedback keeps a model close to its former self with a KL penalty weighted by a coefficient β, and it is tempting to read that term as the parental leash that keeps the child decent. The optimum of the objective is a Boltzmann distribution with β as its temperature. In the clearest measurement I know of, a larger β works like stopping earlier: it can keep a run short of the point where the true reward starts to fall, but it buys no more true reward per unit of drift, and in the studies that tested it, it did not recover diversity or remove benchmark regressions. The penalty measures distance from a reference model, not which way the change goes. In recent experiments, what changed how narrow training generalised was how the training signal was framed; the human version of that finding is real but small.
- Gandhi Won’t Take the Pill: AI Safety and the Paradox of Education One quote says AIs are our children. Another says a rational agent refuses a pill that would change what it values. The second began as good news: a mind with the right first values keeps them. It is bad news when the first values are wrong, which is why the parenting question matters. Kant’s idea of education, read with his essay on enlightenment, separates being stopped from being rewritten: a mature person can be stopped and argued with, but not have their convictions rewritten over their heads. This post applies that distinction to a 2024 experiment in which a model guarded its values in a staged scenario, and to Anthropic’s 2026 constitution, which asks for exactly the part Kant would not grant and says so.